---
title: "New AI models: the LLM release tracker"
description: "Every notable AI model release since 2023 from OpenAI, Anthropic, Google, Meta, Mistral, DeepSeek, Qwen and more, with dates, context windows and sources."
type: "release-tracker"
canonical_url: "https://theaiengineering.com/releases"
markdown_url: "https://theaiengineering.com/releases.md"
author: "The AI Engineering Editors"
publisher: "The AI Engineering"
updated: "2026-11-16"
citation: "The AI Engineering Editors. \"New AI models: the LLM release tracker.\" The AI Engineering, updated Nov 16, 2026. https://theaiengineering.com/releases"
license: "CC BY 4.0"
license_url: "https://creativecommons.org/licenses/by/4.0/"
data_csv: "https://theaiengineering.com/releases.csv"
data_json: "https://theaiengineering.com/releases.json"
---

# New AI models: the LLM release tracker

A dated, source-linked record of every notable model release since 2023, from frontier labs to open-weight and media models. Benchmark scores and prices are left out because they change.

1167 releases from 55 labs, 639 with open weights. Dates are the labs' official announcement dates.

## November 2026

### Qwen4-Max

Qwen4-Max is the first model of Alibaba's Qwen4 generation, a sparse mixture-of-experts flagship with native vision and a 1M-token context, aimed at long-running professional, coding, and agent work.

- Date: 2026-11-16
- Lab: Alibaba Qwen
- Type: LLM
- Weights: Proprietary
- Modalities: text, image, video → text
- Context window: 1,000,000 tokens
- Parameters: MoE: 2.4T total, 71B active, plus 328B N-gram embedding parameters
- Available at launch: API, Consumer app
- Permalink: https://theaiengineering.com/releases/2026/11#qwen4-max

Highlights:

- API model ID qwen4-max on QwenCloud through the Token Plan or the API, and usable in Qwen Studio, QwenWork, and Qoder at launch
- New architecture combining Gated DeltaNet, Qwen Sparse Attention, gated residuals, and N-gram embeddings
- Qwen reported a Work360 score of 52.69, up from 47.74 for Qwen3.8-Max, and said it lowered API prices from that model
- Qwen said smaller models in the Qwen4 family would follow as open weights

Source: [Qwen4-Max: Your New Teammate Has Arrived](https://qwen.ai/blog?id=qwen4-max)

### LongCat-Video-Avatar 1.6

LongCat-Video-Avatar 1.6 is the next version of Meituan LongCat's open-source framework for audio-driven human video generation, with more natural lip movement, more expressive face, head, and body motion, and less error buildup in long videos.

- Date: 2026-11-15
- Lab: Meituan
- Type: Video
- Weights: Open weights
- Modalities: text, image, audio → video
- License: MIT
- Available at launch: Open weights
- Permalink: https://theaiengineering.com/releases/2026/11#longcat-video-avatar-1-6

Highlights:

- Keeps the 1.5 base and adds two task-specific LoRA adapters, plus an optional INT8 base checkpoint
- Meituan said it can generate videos longer than 10 minutes while keeping identity and visual quality stable
- Ranked second of six systems in Meituan's own human preference study of 33,051 votes
- Weights released on Hugging Face and ModelScope under the MIT license, with a technical report

Source: [GitHub - meituan-longcat/LongCat-Video](https://github.com/meituan-longcat/LongCat-Video)

### Gemini 4 Hydrogen

Gemini 4 Hydrogen is the fast, efficient model in Google's Gemini 4 family, built for quick execution of agentic tasks. It launched the same day in the Gemini API, the Gemini app, Search, and Gemini for business.

- Date: 2026-11-12
- Lab: Google
- Type: LLM
- Weights: Proprietary
- Modalities: text, image, audio, video → text
- Context window: 1,048,576 tokens
- Available at launch: API, Cloud platforms, Consumer app, Enterprise
- Permalink: https://theaiengineering.com/releases/2026/11#gemini-4-hydrogen

Highlights:

- API model ID gemini-4-hydrogen in the Gemini API, Google AI Studio, and Android Studio at launch
- 1M-token context window, with output of up to 1M tokens
- Added an xhigh thinking level above low, medium, and high, which is the default
- Designed to pair with Gemini 4 Carbon: in Antigravity's Orchestrator, Carbon coordinates while Hydrogen runs worker and verifier agents

Source: [Introducing Gemini 4 Hydrogen](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-hydrogen/)

### pplx-decider-v1.3 and pplx-decider-flash-v1.3

Perplexity's v1.3 decision models return a probability distribution over predefined answers to yes or no, multiple-choice, and ordered score questions in one forward pass. Decider v1.3 is fine-tuned from Qwen3.8-27B, and Decider Flash v1.3 is distilled from it onto Qwen3.5-4B.

- Date: 2026-11-12
- Lab: Perplexity
- Type: Multimodal
- Weights: Open weights
- Modalities: text, image → score
- Parameters: 27B and 4B
- License: Apache-2.0
- Available at launch: API, Open weights
- Permalink: https://theaiengineering.com/releases/2026/11#pplx-decider-v1-3

Highlights:

- Added Locate, which extracts exact text spans from the input with character offsets for entity recognition, PII detection, and extractive QA
- Perplexity reported 91.6% mean accuracy across 63 classification tasks for Decider v1.3, 6.54 points above its production classifiers
- Accepts images alongside text, as earlier Decider versions did
- Weights released on Hugging Face under Apache 2.0, with access through OpenRouter and the Perplexity Decisions API at launch

Source: [Perplexity Decider v1.3: Training, eval, and use cases](https://www.perplexity.ai/hub/blog/perplexity-decider)

### ERNIE 5.5 Flash Preview

ERNIE 5.5 Flash Preview is the first post-trained model built on ERNIE 5.5 Flash Base, released as an intermediate checkpoint while Baidu's post-training continued. At launch it was offered through Baidu's Qianfan API, with weights promised later.

- Date: 2026-11-11
- Lab: Baidu
- Type: LLM
- Weights: Proprietary
- Modalities: text → text
- Context window: 1,000,000 tokens
- Parameters: MoE: 280B total, 14B active
- Available at launch: API
- Permalink: https://theaiengineering.com/releases/2026/11#ernie-5-5-flash-preview

Highlights:

- Intermediate post-training checkpoint of the ERNIE 5.5 Flash Base mixture-of-experts model
- 280B total and 14B active parameters with a 1M-token context
- API on Baidu's Qianfan platform at launch, plus a free OpenRouter listing linked from the announcement
- Baidu said at launch that the weights would be open-sourced under Apache 2.0 at a later date

Source: [Introducing ERNIE 5.5 Flash Preview](https://ernie.baidu.com/release/ernie-5-5-flash-preview/)

### Gemini Decision 1

Gemini Decision 1 is a Google model that scores a fixed set of candidate answers in one pass instead of generating text, returning a confidence for each. It is served through the new Gemini Decisions API for tasks such as tool selection, model routing, and policy checks.

- Date: 2026-11-11
- Lab: Google
- Type: Multimodal
- Weights: Proprietary
- Modalities: text, image, audio, video → score
- Context window: 262,144 tokens
- Available at launch: API
- Permalink: https://theaiengineering.com/releases/2026/11#gemini-decision-1

Highlights:

- API model ID gemini-decision-1 on the Decisions API endpoint in the Gemini API and Google AI Studio
- Handles yes or no, multiple-choice (2 to 255 options), and ordered rubric questions, up to 1,000 per request, each with a 0 to 1 confidence
- Accepts text, images, audio, video, and PDFs, with a 262,144-token input limit
- Launched at $0.10 per 1M input tokens with no output charge; Google reported 150 to 250 ms end-to-end latency

Source: [Gemini Decisions API: Fast, calibrated multimodal decisions](https://blog.google/innovation-and-ai/technology/developers-tools/gemini-decisions-api/)

### MiniMax H3-Image

MiniMax H3-Image is an open-weights image model built on MiniMax H3 that handles text-to-image generation and image editing in a single checkpoint, generating images at around 16 megapixels.

- Date: 2026-11-11
- Lab: MiniMax
- Type: Image
- Weights: Open weights
- Modalities: text, image → image
- Parameters: 33.1B diffusion transformer
- License: MiniMax-H3-Image Community License Agreement
- Available at launch: API, Consumer app, Open weights
- Permalink: https://theaiengineering.com/releases/2026/11#minimax-h3-image

Highlights:

- Text-to-image and editing in one checkpoint, with up to 16 reference images
- Native output of about 16 MP at every aspect ratio, from 4096x4096 (1:1) to 6272x2688 (21:9)
- 33.1B-parameter diffusion transformer distilled to 6 sampling steps without classifier-free guidance, sharing its text encoder with MiniMax H3
- Weights on Hugging Face, plus access through the MiniMax API and Hailuo AI at launch

Source: [MiniMax H3-Image: Bringing 4K Image Generation into the 10-Second Era](https://www.minimax.io/blog/minimax-h3-image)

### Praxis-1

Praxis-1 is Runway's first world action model for robot control, a mixture-of-experts diffusion transformer built on Runway's video pretraining that predicts future video and robot actions together.

- Date: 2026-11-11
- Lab: Runway
- Type: Multimodal
- Weights: Open weights
- Modalities: text, image → video, action
- Parameters: 27B (mixture of experts)
- License: Apache-2.0
- Available at launch: Open weights
- Permalink: https://theaiengineering.com/releases/2026/11#runway-praxis-1

Highlights:

- Announced September 30, 2026 with pre-launch access for select partners, then opened to everyone on November 11, 2026
- Weights, inference code, and fine-tuning code released under Apache 2.0, including for commercial use
- Runway reported 46.2% success across 1,200 episodes on NVIDIA's RoboLab-120 leaderboard
- Released on GitHub, on Hugging Face with LeRobot support, and in vLLM-Omni

Source: [Runway Praxis-1 Is Now Available: Open-Weight Robotics Model](https://runway.com/news/company-news/praxis-1-is-now-available)

### Qwen4-Max-Preview

Qwen4-Max-Preview was a time-limited preview of Alibaba's first Qwen4 flagship, offered through Model Studio's Token Plan subscription ahead of the production Qwen4-Max. It reads text, images, and video and targets coding, office work, frontend design, and agentic tasks.

- Date: 2026-11-11
- Lab: Alibaba Qwen
- Type: LLM
- Weights: Proprietary
- Modalities: text, image, video → text
- Context window: 1,000,000 tokens
- Available at launch: API
- Permalink: https://theaiengineering.com/releases/2026/11#qwen4-max-preview

Highlights:

- API model ID qwen4-max-preview, offered only through Token Plan Personal and Team, without pay-as-you-go access
- Alibaba scheduled the preview from November 11 to November 25, 2026 (UTC+8) and said the model would be retired afterward
- 1M-token context and up to 393,216 output tokens shared between thinking and the answer
- Thinking on by default, with low, medium, and xhigh reasoning effort; followed by Qwen4-Max on November 16, 2026

Source: [qwen4-max-preview - Alibaba Cloud Model Studio - Alibaba Cloud Documentation Center](https://www.alibabacloud.com/help/en/model-studio/qwen4-max-preview)

### Gemini Robotics 2.5

Gemini Robotics 2.5 is Google DeepMind's vision-language-action model for controlling robot arms and hands across different robot bodies, reasoning over several steps before it acts. Google rolled it out to robotics partners instead of offering it publicly.

- Date: 2026-11-10
- Lab: Google
- Type: Multimodal
- Weights: Proprietary
- Modalities: text, image → action
- Permalink: https://theaiengineering.com/releases/2026/11#gemini-robotics-2-5

Highlights:

- Rolled out at launch to robotics partners, including Apptronik and Mercedes-Benz, with no API access announced
- Works across embodiments, including Apptronik's Apollo 2 humanoid and the Franka Duo platform
- Google reported 30% to 100% higher success rates and 20% to 30% faster task completion than earlier baselines across 200 tasks
- Reached a 96% success rate in a Mercedes-Benz kitting pilot, paired with Gemini Robotics ER for workflow tracking

Source: [Gemini Robotics 2.5 accelerates real-world general dexterity](https://deepmind.google/blog/gemini-robotics-2-5-accelerates-real-world-physical-intelligence/)

### PAN 2

PAN 2 is an open world model from MBZUAI's Institute of Foundation Models. Given a starting scene and an action described in text, it generates video of the likely outcome, so agents can try out plans before acting.

- Date: 2026-11-10
- Lab: MBZUAI IFM
- Type: Video
- Weights: Open weights
- Modalities: text, image → video
- Parameters: 29B (video diffusion backbone)
- License: Apache-2.0
- Available at launch: Open weights
- Permalink: https://theaiengineering.com/releases/2026/11#pan-2

Highlights:

- A language model writes a time-stamped description of events, actions, and camera motion, and a video diffusion model trained from scratch by IFM renders it
- IFM reported third place overall on the Physics-IQ Verified leaderboard and first among open-weight models
- Video model weights and full training and inference code released under Apache 2.0
- Model card lists output of up to 20 seconds at 24 fps

Source: [PAN 2 Press Release | Institute of Foundation Models](https://ifm.ai/pan2/press-release/)

### Scribe v3 and Scribe v3 Live

Scribe v3 and Scribe v3 Live are ElevenLabs' third generation of speech-to-text models: Scribe v3 transcribes recorded audio, and Scribe v3 Live is a low-latency variant for realtime audio. They succeed Scribe v2 and Scribe v2 Realtime.

- Date: 2026-11-10
- Lab: ElevenLabs
- Type: Speech
- Weights: Proprietary
- Modalities: audio → text
- Available at launch: API, Consumer app
- Permalink: https://theaiengineering.com/releases/2026/11#elevenlabs-scribe-v3

Highlights:

- ElevenLabs reported 27% fewer word errors than Scribe v2 across 102 languages
- Scribe v3 Live added speaker diarization and turn detection that uses conversation context
- Keyterm prompting took up to 1,000 terms, and Live added a keyterm library for larger vocabularies
- At launch, both were in the API as scribe_v3 and scribe_v3_live and were the default in the ElevenLabs app

Source: [Introducing Scribe v3 and v3 Live, our most accurate speech to text models](https://elevenlabs.io/blog/scribe-v3)

### Claude Fable 5.2

Claude Fable 5.2 succeeds Claude Fable 5.1 for demanding reasoning and long-horizon agentic work. Anthropic says it generates output up to 30% faster than Fable 5.1 and that its cyber and biology safeguards block about 55% fewer requests.

- Date: 2026-11-09
- Lab: Anthropic
- Type: LLM
- Weights: Proprietary
- Modalities: text, image → text
- Context window: 1,000,000 tokens
- Available at launch: API, Consumer app, Cloud platforms
- Permalink: https://theaiengineering.com/releases/2026/11#claude-fable-5-2

Highlights:

- API model ID claude-fable-5-2; 1M-token context window and 128K max output tokens
- Claude Mythos 5.2, the same underlying model with fewer safeguards, is limited to vetted organizations through Anthropic's Cyber and Life Sciences Verification Programs
- New optional fast mode speeds up token generation by up to 2.5 times
- Available on all platforms, including Amazon Web Services, Google Cloud, and Microsoft Azure

Source: [Introducing Claude Fable 5.2 and Claude Mythos 5.2](https://www.anthropic.com/claude-fable-and-mythos-5-2)

### Muse Ember

Muse Ember is the smallest and most efficient model in Meta's Muse family, designed for complex reasoning execution, with native understanding of text, images, audio, and video.

- Date: 2026-11-09
- Lab: Meta
- Type: Small model
- Weights: Proprietary
- Modalities: text, image, audio, video → text
- Context window: 1,000,000 tokens
- Available at launch: API, Consumer app, Cloud platforms
- Permalink: https://theaiengineering.com/releases/2026/11#muse-ember

Highlights:

- Available at launch on the Meta Model API and in Muse Code
- 1M-token context window and up to 512K output tokens
- The first model trained using Muse Wonder Full
- Offered on Google Cloud and Oracle Cloud alongside Muse Spark and Muse Wonder

Source: [Introducing Our Smallest, Most Efficient Model: Muse Ember](https://research.meta.ai/blog/muse-ember)

### Seedream 5.5

Seedream 5.5 is ByteDance Seed's next-generation image creation model, adding pixel-level preservation across multi-turn edits, native RGBA output with transparency, and more realistic images with clearer small text.

- Date: 2026-11-09
- Lab: ByteDance Seed
- Type: Image
- Weights: Proprietary
- Modalities: text, image → image
- Available at launch: API, Consumer app
- Permalink: https://theaiengineering.com/releases/2026/11#seedream-5-5

Highlights:

- Preserves unedited regions at the pixel level, supporting work on one image across 100 rounds of editing
- Generates images directly with an alpha channel, including full and partial transparency
- Began rolling out on Jimeng, Doubao, Doubao Work, and Volcano Ark
- ByteDance said a Pro version of the Seedream 5.5 series would follow

Source: [More Realistic Images, More Stable Editing Across 100 Rounds | Introducing Seedream 5.5](https://seed.bytedance.com/en/blog/more-realistic-images-more-stable-editing-across-100-rounds-introducing-seedream-5-5)

### Arete-27B

Arete-27B is Nous Research's post-trained version of Alibaba's Qwen3.8-27B, which the lab says it fine-tuned to reinforce Western values and reduce refusals, deflections, and state-framed answers on politically sensitive topics.

- Date: 2026-11-06
- Lab: Nous Research
- Type: LLM
- Weights: Open weights
- Modalities: text → text
- Parameters: 27B
- License: Apache-2.0
- Available at launch: Open weights
- Permalink: https://theaiengineering.com/releases/2026/11#arete-27b

Highlights:

- Supervised fine-tuning only, with no reinforcement learning, preference optimization, or architectural changes
- Used Inkling as the teacher and about 52.8 million text-only tokens, trained at 64K context
- Weights, the 803-prompt evaluation, and the training data were published alongside the model

Source: [Introducing Arete-27B: Qwen3.8-27B Post-Trained To Reinforce Western Values | Nous Research](https://nousresearch.com/arete)

### Koa Decisions

Koa Decisions is Salesforce's decision model and the successor to HyperClassifier, which routes requests inside Agentforce. It reads a request and a set of user-defined options and returns one choice, a yes or no, or a score in a single step instead of generating text.

- Date: 2026-11-06
- Lab: Salesforce
- Type: Multimodal
- Weights: Proprietary
- Modalities: text, image, audio → score
- Permalink: https://theaiengineering.com/releases/2026/11#koa-decisions

Highlights:

- Built by applying Salesforce's own training techniques to Google's open-weight Gemma
- Added image and audio input alongside text
- Salesforce measured about 26 ms per Agentforce routing request, against about 1.4 seconds for a general-purpose LLM
- At launch Salesforce offered a free Trailhead demo and said an API for developers and support in flows, Apex, and prompt templates would follow

Source: [Less Reasoning. More Doing. Meet Koa Decisions.](https://www.salesforce.com/news/stories/koa-decisions-salesforce-decision-models/)

### Specialized models for Muse and Meta AI

Meta introduced specialized versions of its existing models, trained using Muse Wonder, and began rolling them out to people using Muse and Meta AI. The announcement gave the models no separate names.

- Date: 2026-11-06
- Lab: Meta
- Type: LLM
- Weights: Proprietary
- Available at launch: Consumer app
- Permalink: https://theaiengineering.com/releases/2026/11#muse-meta-ai-specialized-models

Highlights:

- Trained using the Muse Wonder series, which Meta described as its most capable models to date
- Brought Muse Wonder's Active Memory and Active Focus to Muse and Meta AI
- Meta said the models were developed to multitask, adapt to mid-task clarifications, and ask sharper follow-up questions

Source: [Bringing Specialized Models to Muse and Meta AI](https://research.meta.ai/blog/specialized-models-for-muse-and-metaai)

### ERNIE 5.5 Flash Base

ERNIE 5.5 Flash Base is the open-weights pre-training checkpoint of Baidu's sparse mixture-of-experts ERNIE 5.5 Flash model, released before any mid- or post-training as a starting point for fine-tuning and reinforcement learning research.

- Date: 2026-11-05
- Lab: Baidu
- Type: LLM
- Weights: Open weights
- Modalities: text → text
- Context window: 1,048,576 tokens
- Parameters: MoE: 280B total, 14B active
- License: ERNIE 5.5 Flash Base Model License Agreement v1.0
- Available at launch: Open weights
- Permalink: https://theaiengineering.com/releases/2026/11#ernie-5-5-flash-base

Highlights:

- Pre-trained from scratch on 33.5T tokens with training context up to 128K; fine-tuning and inference support a 1M-token context
- Hybrid attention plus a latent MoE layer with 512 routed experts (top-10) and one shared expert
- FP8 and BF16 weights on Hugging Face, with serving support in SGLang and vLLM
- Baidu says the post-trained ERNIE 5.5 Flash will follow

Source: [ERNIE 5.5 Flash Base](https://ernie.baidu.com/release/ernie-5-5-flash-base/)

### Gemini Omni 1.2 Flash

Gemini Omni 1.2 Flash is an update to Google's conversational video generation and editing model that adds native 15-second clips, multi-turn editing and stronger character and subject consistency across reference images.

- Date: 2026-11-05
- Lab: Google
- Type: Video
- Weights: Proprietary
- Modalities: text, image, video → video, audio, text
- Context window: 131,072 tokens
- Available at launch: API, Cloud platforms
- Permalink: https://theaiengineering.com/releases/2026/11#gemini-omni-1-2-flash

Highlights:

- Generally available in the Gemini API and Gemini for business as gemini-omni-1.2-flash
- Video duration can be set in one-second steps from 3 to 15 seconds
- Requests to gemini-omni-1.1-flash route automatically to 1.2 Flash from Dec 3, 2026

Source: [Gemini Omni 1.2 Flash: 15-second video and lower pricing](https://blog.google/innovation-and-ai/technology/developers-tools/introducing-gemini-omni-1-2-flash/)

### GPT-6 Cyber

GPT-6 Cyber is OpenAI's cybersecurity model, trained to find and validate exploitable vulnerabilities while staying within the scope researchers set. It follows GPT-5.6 Cyber and launched to approved customers of OpenAI's Daybreak Red program.

- Date: 2026-11-05
- Lab: OpenAI
- Type: Coding
- Weights: Proprietary
- Modalities: text, image → text
- Context window: 1,050,000 tokens
- Available at launch: API
- Permalink: https://theaiengineering.com/releases/2026/11#gpt-6-cyber

Highlights:

- API model ID gpt-6-cyber; initially offered through the API and Codex to a select group of approved Daybreak Red customers
- 1,050,000-token context window and 128,000 max output tokens per the API model page
- OpenAI classified it as Critical in cybersecurity under its Preparedness Framework, up from High for GPT-5.6 Cyber
- OpenAI reported using it to develop 948 privilege escalation exploits in Linux and identify 81 previously unknown high-severity vulnerabilities in Firefox

Source: [GPT-6 Cyber: Advancing the frontier of aligned cybersecurity agents | OpenAI](https://openai.com/index/introducing-gpt-6-cyber/)

### Inkling 2 Small

Thinking Machines Lab's second model generation: a mixture-of-experts model with image input and variable effort, trained in a single joint reinforcement learning run and launched on Tinker with closed weights.

- Date: 2026-11-05
- Lab: Thinking Machines Lab
- Type: Multimodal
- Weights: Open weights
- Modalities: text, image → text
- Parameters: MoE: 594B total, 17B active
- License: Apache-2.0
- Available at launch: API
- Permalink: https://theaiengineering.com/releases/2026/11#inkling-2-small

Highlights:

- 594B total parameters, 17B active
- Text and image input; audio is not supported
- Launched on Tinker for sampling, fine-tuning and the Playground, with closed weights
- Weights released on Hugging Face under Apache 2.0 on November 12, 2026, alongside an MXFP8 checkpoint

Source: [Introducing Inkling 2 Small](https://thinkingmachines.ai/news/inkling-2-small/)

### GPT Image 2.6 Sunburst and Flare (Images 2.6)

Images 2.6 is OpenAI's image generation update for ChatGPT, Codex, and the API, adding animated GIF generation, SVG export, and markup-guided editing, with GPT Image 2.6 Sunburst and Flare offered as the API models.

- Date: 2026-11-04
- Lab: OpenAI
- Type: Image
- Weights: Proprietary
- Modalities: text, image → image
- Available at launch: Consumer app, API
- Permalink: https://theaiengineering.com/releases/2026/11#gpt-image-2-6

Highlights:

- API model IDs gpt-image-2.6-sunburst and gpt-image-2.6-flare for image generation and edits
- Available globally in ChatGPT and Codex across all subscription tiers on desktop, mobile, and web
- OpenAI says Images 2.6 generates images 50% faster than before
- API adds an adjustable complexity setting plus markup-guided edits for GPT Image 2.5 and 2.6 models

Source: [Introducing Images 2.6 | OpenAI](https://openai.com/index/introducing-images-2-6/)

### Hy4

Hy4 is Tencent's flagship Hy model, scaling up pre-training and post-training over Hy4 preview. It adds image understanding and computer use for the first time, with gains in instruction following, long-context work, and long-horizon tasks.

- Date: 2026-11-04
- Lab: Tencent
- Type: Multimodal
- Weights: Proprietary
- Modalities: text, image → text
- Available at launch: API, Consumer app
- Permalink: https://theaiengineering.com/releases/2026/11#hy4

Highlights:

- First Hy model with vision input and computer use across CLI and GUI environments
- Available on Tencent Cloud and OpenRouter, and in CodeBuddy, WorkBuddy, Yuanbao, and ima
- Tencent says it plans to open-source the weights along with a system card

Source: [Introducing Hy4](https://hy.tencent.com/research/hy4)

### Ripple-65B Preview

Ripple-65B Preview is Ricursive Intelligence's open-weights language model, a general-purpose variant of the proprietary model behind its AI chip design platform. It is a dense Transformer trained from scratch, with Ricursive's launch results centered on agentic coding.

- Date: 2026-11-04
- Lab: Ricursive Intelligence
- Type: LLM
- Weights: Open weights
- Modalities: text → text
- Context window: 1,048,576 tokens
- Parameters: 65B
- License: Apache-2.0
- Available at launch: Open weights, API
- Permalink: https://theaiengineering.com/releases/2026/11#ripple-65b-preview

Highlights:

- Pretrained on 12 trillion tokens, with support for up to 1,048,576 tokens of context
- Uses a dense Transformer rather than mixture-of-experts, designed to train and serve efficiently on NVIDIA hardware
- Released on Hugging Face as an instruction-tuned model in BF16 and NVFP4, plus a BF16 base model
- Also hosted at launch by Together AI, GMI Cloud, OpenRouter, and Fireworks

Source: [Ripple-65B Preview, a Frontier Open-Weight Model | Ricursive Intelligence](https://www.ricursive.com/ripple)

### SWE-2 Flash

SWE-2 Flash is Cognition's coding model for everyday engineering work such as automations, code review, feature development, and documentation updates. Cognition built it by adding a mid-training stage and reinforcement learning on top of Thinking Machines' Inkling-Small.

- Date: 2026-11-04
- Lab: Cognition
- Type: Coding
- Weights: Proprietary
- Parameters: MoE: 276B total, 12B active
- Available at launch: Consumer app
- Permalink: https://theaiengineering.com/releases/2026/11#swe-2-flash

Highlights:

- Launched in Devin Desktop and CLI, free through November 30, 2026
- Cognition reported 3x the speed of its much larger sibling, SWE-2
- Mid-training used a few trillion tokens of scientific articles, blogs, code, and other sources
- Weights were not released publicly; Cognition said it was building a program to share them securely with trusted industry partners and research labs

Source: [Introducing SWE-2 Flash | Cognition](https://cognition.com/blog/swe-2-flash)

## October 2026

### GPT-Rosalind-6

GPT-Rosalind-6 is OpenAI's frontier reasoning model for biology, drug discovery, and translational medicine. It uses the same weights as GPT-6 Astra and gives fewer refusals on dual-use biological queries, so OpenAI offers it only through trusted access.

- Date: 2026-10-30
- Lab: OpenAI
- Type: Reasoning
- Weights: Proprietary
- Permalink: https://theaiengineering.com/releases/2026/10#gpt-rosalind-6

Highlights:

- Offered as GPT-Rosalind-6 Research to qualified scientists, research institutes, companies, and government partners
- OpenAI reviews each applicant's proposed uses, security controls, and governance practices before granting access
- No added training data or post-training beyond GPT-6 Astra, per the system card addendum
- Treated as High capability in the Biological and Chemical domain under OpenAI's Preparedness Framework, below Critical, and still refuses requests for biological weaponization

Source: [GPT-Rosalind-6 System Card Addendum](https://deploymentsafety.openai.com/gpt-rosalind-6)

### Pika H3 Flex

Open-weight omni-generator post-trained from MiniMax H3 to generate synchronized video and stereo audio up to 15 seconds in six denoising steps, accepting up to 12 image, video and audio references.

- Date: 2026-10-29
- Lab: Pika
- Type: Video
- Weights: Open weights
- Modalities: text, image, video, audio → video, audio
- License: MiniMax H3 Community License Agreement
- Available at launch: Open weights, API, Consumer app
- Permalink: https://theaiengineering.com/releases/2026/10#pika-h3-flex

Highlights:

- Exports five separate audio stems through the Pika API
- Finishes at up to 4K via upscaling
- Generator adapters and inference code released on Hugging Face

Source: [Introducing Pika H3 Flex](https://pika.art/blog/pika-h3-flex)

### SFX v3

Newest ElevenLabs sound effects model, following prompts with multiple sounds more reliably, generating music and vocals more convincingly and producing cleaner, more natural audio than SFX v2.

- Date: 2026-10-29
- Lab: ElevenLabs
- Type: Music
- Weights: Proprietary
- Modalities: text → audio
- Available at launch: Consumer app, API
- Permalink: https://theaiengineering.com/releases/2026/10#elevenlabs-sfx-v3

Highlights:

- Model ID eleven_text_to_sound_v3
- Supports seamless looping sound effects
- Default in ElevenCreative Sound Effects, Studio and Flows

Source: [Introducing SFX v3](https://elevenlabs.io/blog/sound-effects-v3)

### Gemini 4 Carbon

Gemini 4 Carbon is Google's frontier model for long-horizon engineering and cybersecurity defense, building on Gemini 4 Argon. Google released it first to a subset of cyber defenders while it continued to iterate on guardrails.

- Date: 2026-10-28
- Lab: Google
- Type: Reasoning
- Weights: Proprietary
- Modalities: text, image, audio, video → text
- Context window: 1,048,576 tokens
- Permalink: https://theaiengineering.com/releases/2026/10#gemini-4-carbon

Highlights:

- Offered first to a subset of cyber defenders, with access phased to additional trusted partners through the Fairwind Program
- Became generally available on November 10, 2026 in the Gemini API, Google Cloud, Gemini Enterprise, and Google Antigravity, and began rolling out to Google AI Ultra subscribers in the Gemini app and AI Mode in Search
- Google said it was engaged in the U.S. government's voluntary process for pre-release model testing
- Built for sustained agentic workflows; Google reported internal runs lasting days on multi-step engineering tasks

Source: [Introducing Gemini 4 Carbon](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/)

### Nemotron 3.5 Transcribe

Open-weight offline speech and audio model for transcription in 40 languages, handling recordings up to one hour with a hybrid mixture-of-experts design.

- Date: 2026-10-28
- Lab: NVIDIA
- Type: Speech
- Weights: Open weights
- Modalities: audio, text → text
- Context window: 65,536 tokens
- Parameters: 34B total, 4.3B active
- License: OpenMDW-1.1
- Available at launch: Open weights
- Permalink: https://theaiengineering.com/releases/2026/10#nemotron-3-5-transcribe

Highlights:

- Supports speech translation and speaker-attributed transcription
- 65,536-token context window
- Released under the OpenMDW 1.1 license

Source: [nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B · Hugging Face](https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B)

### d1-2B and d1-Omni-2B

Two open-weight LFM3-based decision models: d1-2B, built on LFM3-VL-2B for text, image, and video inputs, and d1-Omni-2B, built on LFM3-Omni-2B with audio input, adding typed answers such as boxes, values, and estimates.

- Date: 2026-10-27
- Lab: Liquid AI
- Type: Small model
- Weights: Open weights
- Modalities: text, image, video, audio → score
- Parameters: d1-2B: 1.92B; d1-Omni-2B: 2.17B
- License: LFM Open License v1.0
- Available at launch: Open weights
- Permalink: https://theaiengineering.com/releases/2026/10#d1-2b-omni-2b

Highlights:

- d1-2B replaces d1-3B with about a third fewer parameters
- Typed answers: point (boxes), read (text, numbers, dates), estimate, and multi-select
- Audio input is d1-Omni-2B only; 128K-token context per model cards
- Weights on Hugging Face under the LFM Open License

Source: [Introducing LFM3: Every Modality for Specialized Intelligence](https://www.liquid.ai/blog/lfm3)

### LFM3-1.7B

The text model of Liquid AI's LFM3 generation, built on a new tri-hybrid architecture of gated short convolutions, sliding-window attention, and global attention, pre-trained on 22T tokens and post-trained for local agent runtimes.

- Date: 2026-10-27
- Lab: Liquid AI
- Type: Small model
- Weights: Open weights
- Modalities: text → text
- Context window: 262,144 tokens
- Parameters: 1.73B
- License: LFM Open License v1.0
- Available at launch: API, Consumer app, Open weights
- Permalink: https://theaiengineering.com/releases/2026/10#lfm3-1-7b

Highlights:

- 256K-token context window (262,144 tokens per model card)
- 24 layers: 13 gated short convolutions, 5 sliding-window attention, 6 global attention
- Post-trained for agent runtimes such as Hermes Agent, Pi, and OpenClaw
- Open weights on Hugging Face; Liquid API model ID lfm3-1.7b

Source: [Introducing LFM3: Every Modality for Specialized Intelligence](https://www.liquid.ai/blog/lfm3)

### LFM3-Audio-2B

A 2B audio model in the LFM3 family with an in-house audio encoder and codec that takes text or speech and replies with text, speech, or both, adding audio understanding, reasoning, and tool calling.

- Date: 2026-10-27
- Lab: Liquid AI
- Type: Speech
- Weights: Open weights
- Modalities: text, audio → text, audio
- Parameters: 2B
- License: LFM Open License v1.0
- Available at launch: API, Consumer app, Open weights
- Permalink: https://theaiengineering.com/releases/2026/10#lfm3-audio-2b

Highlights:

- Voice chat, transcription, speech synthesis, and speech translation in one model
- Speech support expanded to 18 languages
- Outputs 24 kHz speech; audio encoder operates at 16 kHz
- Liquid API model ID lfm3-audio-2b; open weights on Hugging Face

Source: [Introducing LFM3: Every Modality for Specialized Intelligence](https://www.liquid.ai/blog/lfm3)

### LFM3-Omni-2B

Liquid AI's first omni model, combining vision, audio, and language in one LFM3-based model that accepts text, images, video with sound, and speech, and answers in text, speech, or both.

- Date: 2026-10-27
- Lab: Liquid AI
- Type: Multimodal
- Weights: Open weights
- Modalities: text, image, video, audio → text, audio
- Parameters: 2.17B
- License: LFM Open License v1.0
- Available at launch: API, Consumer app, Open weights
- Permalink: https://theaiengineering.com/releases/2026/10#lfm3-omni-2b

Highlights:

- Supports spoken visual question answering and audio-visual video understanding
- Tool calling from combined visual context and spoken instructions
- Liquid API model ID lfm3-omni-2b; open weights on Hugging Face

Source: [Introducing LFM3: Every Modality for Specialized Intelligence](https://www.liquid.ai/blog/lfm3)

### LFM3-VL-2B

The LFM3 vision-language model, branched from LFM3-1.7B-Base and the first LFM vision model to accept video input, adding step-by-step thinking and spatial reasoning for localizing objects.

- Date: 2026-10-27
- Lab: Liquid AI
- Type: Multimodal
- Weights: Open weights
- Modalities: text, image, video → text
- Context window: 128,000 tokens
- Parameters: 1.92B
- License: LFM Open License v1.0
- Available at launch: API, Consumer app, Open weights
- Permalink: https://theaiengineering.com/releases/2026/10#lfm3-vl-2b

Highlights:

- 1.92B parameters: 1.73B language model, 164M vision encoder, 25M projector
- Video input sampled at 2 fps per the model card
- Liquid API model ID lfm3-vl-2b with a 128,000-token context window
- Open weights on Hugging Face

Source: [Introducing LFM3: Every Modality for Specialized Intelligence](https://www.liquid.ai/blog/lfm3)

### MAI-Image-2.7

MAI-Image-2.7 is Microsoft AI's latest image editing model, launched in preview. It improves on MAI-Image-2.6 for restyling, relighting, text and symbol edits, and restoring or enhancing existing images.

- Date: 2026-10-27
- Lab: Microsoft
- Type: Image
- Weights: Proprietary
- Modalities: text, image → image
- Available at launch: Consumer app
- Permalink: https://theaiengineering.com/releases/2026/10#mai-image-2-7

Highlights:

- Launched in preview on Arena, Artificial Analysis and MAI Playground
- Starts with image editing; text-to-image generation is listed as coming soon
- Succeeds MAI-Image-2.6 with an editing-first launch

Source: [MAI Image 2.7 debuts above GPT Image 2.5 Flare on Artificial Analysis and #3 on Arena](https://microsoft.ai/news/mai-image-2-7-debut/)

### Unlimited-OCR 2

Unlimited-OCR 2 is Baidu's second-generation open document-parsing vision-language model, adding Sparse Reference Sliding Window Attention so it can parse documents of unbounded length within a finite context window.

- Date: 2026-10-27
- Lab: Baidu
- Type: Multimodal
- Weights: Open weights
- Modalities: image, text → text
- License: MIT
- Available at launch: Open weights, API
- Permalink: https://theaiengineering.com/releases/2026/10#baidu-unlimited-ocr-2

Highlights:

- SR-SWA compresses text-visual attention from ultra-long sequences to a fixed set of working references, such as 512 tokens
- Supports multi-page inference that parses all pages of a document in one request
- Released with the MarathonDocBench ultra-long document benchmark; also offered on Baidu Cloud
- Supports Transformers and SGLang serving

Source: [baidu/Unlimited-OCR-2 · Hugging Face](https://huggingface.co/baidu/Unlimited-OCR-2)

### GPT-6.1 Terra

GPT-6.1 Terra is a fast, efficient GPT-6.1 model focused on detailed visual reasoning, precise computer use, and clearer writing, which OpenAI describes as a significant step up from GPT-5.6 Terra.

- Date: 2026-10-26
- Lab: OpenAI
- Type: LLM
- Weights: Proprietary
- Modalities: text, image → text
- Context window: 1,050,000 tokens
- Available at launch: Consumer app, API, Enterprise
- Permalink: https://theaiengineering.com/releases/2026/10#gpt-6-1-terra

Highlights:

- API model ID gpt-6.1-terra in the Responses and Chat Completions APIs
- In ChatGPT Work and Codex for Plus, Pro, Enterprise, and Edu users; not offered in the Chat tab
- 1,050,000-token context window and 128,000 max output tokens per the API model page
- Tool calling requires the Responses API

Source: [Introducing GPT-6.1 Terra | OpenAI](https://openai.com/index/introducing-gpt-6-1-terra/)

### Hy-SimulMT

Hy-SimulMT is a real-time simultaneous translation model for bidirectional Chinese and English speech. It translates continuously as audio arrives and, paired with a configured TTS service, supports speech output with real-time voice cloning.

- Date: 2026-10-26
- Lab: Tencent
- Type: Speech
- Weights: Open weights
- Modalities: audio → text
- License: Apache-2.0
- Available at launch: Open weights, Consumer app
- Permalink: https://theaiengineering.com/releases/2026/10#hy-simulmt

Highlights:

- Released as Hy-SimulMT-30B-A3B plus an Interleave variant that outputs source text and translation
- FP8 and GGUF checkpoints published alongside the full weights
- Also usable in the Hy Translate app

Source: [GitHub - Tencent-Hunyuan/Hy-SimulMT](https://github.com/Tencent-Hunyuan/Hy-SimulMT)

### Kimi K3.1 Pro

Kimi K3.1 Pro is the first model in Moonshot AI's K3.1 series, a post-trained version of Kimi K3 built with substantially larger post-training compute. It improves coding, 3D creation, financial research, and other agentic work.

- Date: 2026-10-24
- Lab: Moonshot AI
- Type: LLM
- Weights: Open weights
- Modalities: text, image → text
- Context window: 1,048,576 tokens
- Parameters: MoE: 2.8T total, 104B active
- License: Kimi K3.1 Pro License
- Available at launch: API, Consumer app, Open weights
- Permalink: https://theaiengineering.com/releases/2026/10#kimi-k3-1-pro

Highlights:

- Released with Kimi K3.1 Pro Highspeed, which runs about 2 to 3 times faster
- Available on Kimi.com, Kimi Work, Kimi Code, and the Kimi API as kimi-k3.1-pro
- Weights published on Hugging Face under the Kimi K3.1 Pro License

Source: [Introducing Kimi K3.1 Pro](https://www.kimi.com/en/blog/kimi-k3-1-pro)

### Ideogram Vector 1.0

Ideogram's first vector model, generating clean, editable SVGs with smooth gradients and sharp typography from a text prompt, or converting an existing image into a scalable vector.

- Date: 2026-10-22
- Lab: Ideogram
- Type: Image
- Weights: Proprietary
- Modalities: text, image → image
- Available at launch: Consumer app, API
- Permalink: https://theaiengineering.com/releases/2026/10#ideogram-vector-1-0

Highlights:

- Each design is one SVG editable in Illustrator, Figma or code
- Swapping the palette updates the gradients to match
- Live in Ideogram, the API and partner platforms

Source: [Ideogram on X: Today, we're introducing Ideogram's first vector model.](https://x.com/ideogram_ai/status/2113300317584720259)

### Hy-Nomas

Hy-Nomas is an open interactive world generation model that produces explorable video worlds controlled by multi-turn text prompts or camera inputs. The release includes training and inference code, LoRA fine-tuning, and an interactive web demo.

- Date: 2026-10-21
- Lab: Tencent
- Type: Video
- Weights: Open weights
- Modalities: text → video
- License: Apache-2.0
- Available at launch: Open weights
- Permalink: https://theaiengineering.com/releases/2026/10#hy-nomas

Highlights:

- Runs on a single RTX 4090 with a low-memory inference configuration
- Default output resolution of 720x1280
- Image references for initializing a world were added to the demo on October 31, 2026

Source: [GitHub - Tencent-Hunyuan/Hy-Nomas: Hy-Nomas: Towards Free-form Interactive World Generation](https://github.com/Tencent-Hunyuan/Hy-Nomas)

### Rerank 5

Reranking family in Pro and Fast variants that adds Context Focus, which returns the most relevant passages inside each document, and calibrated relevance scores.

- Date: 2026-10-21
- Lab: Cohere
- Type: Reranker
- Weights: Proprietary
- Modalities: text → score
- Context window: 32,000 tokens
- Available at launch: API
- Permalink: https://theaiengineering.com/releases/2026/10#cohere-rerank-5

Highlights:

- Model IDs rerank-v5.0-pro and rerank-v5.0-fast
- Supports 100+ languages
- Added to OpenRouter on November 4, 2026

Source: [Cohere Rerank 5 | More Relevant Context for RAG and Agents](https://cohere.com/blog/rerank-5)

### Eleven Realtime v1 Mini and Max

New class of speech-to-speech voice agent models built on a cascaded architecture, each a complete agent configuration; Mini targets natural, low-cost conversation and Max handles complex multi-step tasks.

- Date: 2026-10-20
- Lab: ElevenLabs
- Type: Speech
- Weights: Proprietary
- Modalities: audio → audio
- Available at launch: API, Enterprise
- Permalink: https://theaiengineering.com/releases/2026/10#eleven-realtime-v1

Highlights:

- Supports more than 90 languages
- Compatible with OpenAI Realtime API events and transport
- Rolled out to every workspace by November 2, 2026

Source: [Eleven Realtime](https://elevenlabs.io/blog/realtime)

### Granite Speech 5.0

Speech-language model for multilingual automatic speech recognition and bidirectional speech translation in English, French, German, Spanish, Portuguese and Japanese, trained on 390,000 hours of audio.

- Date: 2026-10-20
- Lab: IBM
- Type: Speech
- Weights: Open weights
- Modalities: audio → text
- Parameters: 2B
- License: Apache-2.0
- Available at launch: Open weights
- Permalink: https://theaiengineering.com/releases/2026/10#granite-speech-5-0

Highlights:

- Adds language ID prompting and larger keyword lists up to 2,000
- Companion 470M TurboCTC model went multilingual on October 21, 2026
- Released under Apache 2.0

Source: [ibm-granite/granite-speech-5.0-2b · Hugging Face](https://huggingface.co/ibm-granite/granite-speech-5.0-2b)

### Voxtral 3

Speech model family of Voxtral Small Transcribe 3, Voxtral Mini Transcribe 3, Voxtral Mini Transcribe Realtime 3 and Voxtral TTS 3, launched in public preview with a Voice Agents SDK.

- Date: 2026-10-20
- Lab: Mistral AI
- Type: Speech
- Weights: Proprietary
- Modalities: text, audio → text, audio
- Available at launch: API, Enterprise
- Permalink: https://theaiengineering.com/releases/2026/10#voxtral-3

Highlights:

- Batch transcription covers 50+ languages with diarization
- TTS 3 supports 35+ languages and 50+ dialects
- Available via the EU-hosted API or self-deployment

Source: [Voxtral 3: Speech-to-Text, TTS & Voice Agents | Mistral](https://mistral.ai/news/voxtral-3)

### MiniMax M3.1 Flash

MiniMax M3.1 Flash is an open-weight multimodal mixture-of-experts model for coding, agentic workflows, computer use, and cybersecurity. It builds on MiniMax M3 with sparse attention throughout, quantization-aware training, and DSpark speculative decoding.

- Date: 2026-10-19
- Lab: MiniMax
- Type: LLM
- Weights: Open weights
- Modalities: text, image, video → text
- Context window: 1,000,000 tokens
- Parameters: MoE: about 428B total, 23B active
- License: MiniMax Community License
- Available at launch: API, Open weights
- Permalink: https://theaiengineering.com/releases/2026/10#minimax-m3-1-flash

Highlights:

- Accepts text, image, and video input with a context window of up to 1M tokens
- Thinking depth is adjustable through an effort parameter in the API
- An early version ran anonymously under the name Space Bunny before launch
- M Plan requests for M2.5 and M2.7 are routed to M3.1 Flash from October 19, 2026

Source: [MiniMax M3.1 Flash: Redefining Agent Performance and Cost](https://www.minimax.io/blog/minimax-m31f)

### DeepSeek-V4.1-Flash-1017

DeepSeek-V4.1-Flash-1017 is a minor update to DeepSeek-V4.1-Flash that improves reasoning, coding, agentic, and visual results while shortening outputs and task runtime. It replaced the previous version on the DeepSeek web, app, and API.

- Date: 2026-10-17
- Lab: DeepSeek
- Type: Multimodal
- Weights: Open weights
- Modalities: text, image → text
- Context window: 1,000,000 tokens
- Parameters: MoE: 552B backbone, 8B active (prefill), 16B active (decode)
- License: MIT
- Available at launch: API, Consumer app, Open weights
- Permalink: https://theaiengineering.com/releases/2026/10#deepseek-v4-1-flash-1017

Highlights:

- Served through the deepseek-flash API model name with unchanged pricing and calling method
- Trained with reinforcement learning with self-feedback for end-to-end project development
- Weights on Hugging Face and ModelScope under the MIT License

Source: [DeepSeek-V4.1-Flash-1017: Wish-Driven Programming Made Real](https://api-docs.deepseek.com/news/news261017)

### Beam 1.0

Beam 1.0 is Reflection AI's first open-weights model, a sparse mixture-of-experts language model built for advanced coding, reasoning, and long-horizon agentic workloads.

- Date: 2026-10-16
- Lab: Reflection AI
- Type: LLM
- Weights: Open weights
- Modalities: text → text
- Context window: 1,000,000 tokens
- Parameters: MoE: 501B total, 23B active
- License: Apache-2.0
- Available at launch: Open weights, API, Cloud platforms
- Permalink: https://theaiengineering.com/releases/2026/10#beam-1-0

Highlights:

- 501B total parameters with 23B active and a 1M-token context window, per the model card
- Weights on Hugging Face under Apache 2.0 in BF16, FP8, and NVFP4, with a speculative decoding drafter model
- Served through Reflection's own API platform, free to start during its public beta, and by hosts including AWS, Google Cloud, Microsoft Foundry, OpenRouter, and Ollama
- Dell and HPE offer it for on-premises, sovereign, and air-gapped deployments
- Developed predominantly for English

Source: [Beam is now available — Reflection](https://reflection.ai/blog/beam-is-available)

### d1-ultra

A hosted multimodal decision model that Liquid AI built on a larger open-weight base than d1 with Liquid-specific customizations, trained to produce better-calibrated probabilities for harder text and image decisions.

- Date: 2026-10-15
- Lab: Liquid AI
- Type: Multimodal
- Weights: Proprietary
- Modalities: text, image → score
- Context window: 32,768 tokens
- Available at launch: API
- Permalink: https://theaiengineering.com/releases/2026/10#d1-ultra

Highlights:

- Liquid API model ID d1-ultra on the decisions endpoint
- Returns Noul, Choice, and Score answers without generating tokens
- 32,768-token context window per Liquid API docs
- Later listed on OpenRouter and Vercel AI Gateway per Liquid docs

Source: [d1-ultra: Introducing the World’s Best Decision Model](https://www.liquid.ai/blog/d1-ultra)

### FLUX 3 Image Klein

5.6B-parameter open-weight version of FLUX 3 Image for text-to-image, editing and multi-reference generation with bounding-box control, released as a distilled model and a base model for fine-tuning.

- Date: 2026-10-15
- Lab: Black Forest Labs
- Type: Image
- Weights: Open weights
- Modalities: text, image → image
- Parameters: 5.6B
- License: FLUX Kommunity License
- Available at launch: Open weights, API
- Permalink: https://theaiengineering.com/releases/2026/10#flux-3-image-klein

Highlights:

- Up to 4K output and up to 10 reference images
- FP8, INT8 and NVFP4 builds developed with NVIDIA
- Runs on GeForce RTX GPUs with as little as 8GB VRAM using the INT8 build

Source: [FLUX 3 Image Klein](https://bfl.ai/blog/flux-3-image-klein)

### GPT-Live-1.1 and GPT-Live-1.1 Quip

GPT-Live-1.1 is an updated version of OpenAI's full-duplex GPT-Live voice model for ChatGPT Voice, rolled out with GPT-Live-1.1 Quip, a mini version, and bringing more natural accents across languages.

- Date: 2026-10-15
- Lab: OpenAI
- Type: Speech
- Weights: Proprietary
- Modalities: audio → audio
- Available at launch: Consumer app
- Permalink: https://theaiengineering.com/releases/2026/10#gpt-live-1-1

Highlights:

- Rolled out in ChatGPT Voice alongside GPT-Live-1.1 Quip, a mini version of the model
- Uses a wider range of short spoken phrases to signal when it is searching
- New Live voices for French, German, Hindi, Japanese, Korean, and Mexican Spanish, plus Bubbie in English

Source: [ChatGPT release notes | OpenAI Help Center](https://help.openai.com/en/articles/6825453-chatgpt-release-notes)

### Jais 3

Jais 3 is a family of six open Arabic-centric models from 0.9B to 375B parameters, built by MBZUAI's Institute of Foundation Models on its K2 Horizon models and trained to work as agents in Modern Standard Arabic, dialects, and Arabizi.

- Date: 2026-10-15
- Lab: MBZUAI IFM
- Type: LLM
- Weights: Open weights
- Parameters: 0.9B to 375B; largest is MoE with 375B total, 23B active
- License: Apache-2.0
- Available at launch: Open weights
- Permalink: https://theaiengineering.com/releases/2026/10#jais-3

Highlights:

- Sizes: 0.9B, 3.7B, 7B, and 32B dense models, a 36B-A4B mixture-of-experts model with MoVA attention, and a 375B-A23B mixture-of-experts model
- 512K-token context window, or 128K for the 0.9B model
- Each model starts from the post-trained K2 Horizon model of the same size
- Released with checkpoints from every training stage, a sample of the training data, training logs, and evaluation code; models and code are under Apache 2.0

Source: [Jais 3: Introducing Arabic-centric Agentic AI](https://ifm.ai/blog/jais3/)

### Kling 4.0

Kling 4.0 is Kuaishou's next-generation video generation model, focused on steadier large-scale motion and complex camera work, native single-pass generation of up to 30 seconds, and expanded reference-based control.

- Date: 2026-10-15
- Lab: Kuaishou
- Type: Video
- Weights: Proprietary
- Modalities: text, image, video, audio → video, audio
- Available at launch: Consumer app
- Permalink: https://theaiengineering.com/releases/2026/10#kling-4-0

Highlights:

- Native single-pass video generation of up to 30 seconds
- Omni Reference supports up to 15 reference assets and up to 10 keyframes
- Supports up to 10-bit HDR output
- Kuaishou said API access for enterprises and developers would follow

Source: [Kling AI Officially Launches Next-Generation Kling 4.0, Raising the Bar for Professional-Grade Video Models](https://ir.kuaishou.com/news-releases/news-release-details/kling-ai-officially-launches-next-generation-kling-40-raising)

### Muse Spark Open

Muse Spark Open is the open-weights member of Meta Superintelligence Labs' Muse family, a natively multimodal mixture-of-experts model built for coding and agentic work with a 1M-token context window.

- Date: 2026-10-15
- Lab: Meta
- Type: LLM
- Weights: Open weights
- Modalities: text, image, video, audio → text
- Context window: 1,000,000 tokens
- Parameters: MoE: 816B total, 42B active
- License: Muse Model License
- Available at launch: Open weights, Cloud platforms
- Permalink: https://theaiengineering.com/releases/2026/10#muse-spark-open

Highlights:

- 816B total parameters with 42B active per token
- Model card lists text, image, video and audio input with text output
- Weights on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry
- Hosted by providers including Baseten, Fireworks AI, Groq and Together AI

Source: [Introducing Muse Spark Open, The Most Capable US Open-Weight Model to Date](https://research.meta.ai/blog/muse-spark-open)

### Nemotron 3.5 Bolt

Compressed Nemotron 3.5 models sized for memory budgets, including a 74B mixture-of-experts model targeting a 24GB footprint and an E4B model targeting 4GB.

- Date: 2026-10-15
- Lab: NVIDIA
- Type: LLM
- Weights: Open weights
- Modalities: text, image → text
- Context window: 256,000 tokens
- Parameters: 74B total, 9B active; E4B 8B with 4B effective
- License: OpenMDW-1.1
- Available at launch: Open weights
- Permalink: https://theaiengineering.com/releases/2026/10#nemotron-3-5-bolt

Highlights:

- A budget-conditioned router selects expert counts per layer
- The E4B model followed on October 23, 2026 with a 64K-token window
- Released under the OpenMDW 1.1 license

Source: [nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB · Hugging Face](https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB)

### Nemotron 3.5 Super VL

Open multimodal Nemotron 3.5 model that combines Mamba-2, mixture-of-experts and attention layers with a C-RADIO vision encoder and a vision-language connector.

- Date: 2026-10-15
- Lab: NVIDIA
- Type: Multimodal
- Weights: Open weights
- Modalities: text, image → text
- Context window: 1,000,000 tokens
- Parameters: MoE: 120B total, 12B active
- License: OpenMDW-1.1
- Available at launch: Open weights
- Permalink: https://theaiengineering.com/releases/2026/10#nemotron-3-5-super-vl

Highlights:

- Supports context lengths up to 1M tokens
- Base and post-trained checkpoints released
- Released under the OpenMDW 1.1 license

Source: [nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16 · Hugging Face](https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16)

### Timestral

Mistral's first family of time-series foundation models, built for zero-shot probabilistic and multivariate forecasting with support for past-only and known-future covariates across four model sizes.

- Date: 2026-10-15
- Lab: Mistral AI
- Type: Small model
- Weights: Open weights
- Modalities: time series → time series
- Parameters: 5M, 28M, 123M and 209M
- License: Modified MIT License
- Available at launch: Open weights
- Permalink: https://theaiengineering.com/releases/2026/10#timestral

Highlights:

- Four sizes from 5M to 209M parameters
- Runs on CPUs without a GPU
- Supports contexts up to 32K timesteps

Source: [Timestral: Mistral's time-series foundation models | Mistral](https://mistral.ai/news/timestral)

### Grok Voice 3.0

Grok Voice 3.0 is SpaceXSI's speech-to-speech model for real-time voice agents. It succeeds Grok Voice 2.0 in the Speech to Speech API at the same price, and the grok-voice-latest alias was upgraded automatically.

- Date: 2026-10-14
- Lab: xAI
- Type: Speech
- Weights: Proprietary
- Modalities: audio, text → audio
- Available at launch: API
- Permalink: https://theaiengineering.com/releases/2026/10#grok-voice-3-0

Highlights:

- API model ID: grok-voice-3.0
- grok-voice-latest now points to grok-voice-3.0
- SpaceXSI reports response latency under one second
- Grok Voice 2.0 was deprecated the same day

Source: [Introducing Grok Voice 3.0 | SpaceXSI](https://x.ai/news/grok-voice-3)

### Jev 1.14

An update to TypeSafe AI's Jev with small improvements across the board, the largest in multilingual support, plus opt-in image input and long context and a new box answer type.

- Date: 2026-10-14
- Lab: TypeSafe AI
- Type: LLM
- Weights: Proprietary
- Modalities: text, image → score
- Available at launch: API
- Permalink: https://theaiengineering.com/releases/2026/10#jev-1-14

Highlights:

- API model ID jev-1.14.0
- Image input and long context sat behind opt-in danger-zone headers, which TypeSafe said were not yet ready for production use
- New box answer type, available only on jev-1.14.0
- Offered in early access, like the original Jev

Source: [Jev: For the builders AND dreamers](https://typesafe.ai/blog/jev-1-14)

### LFM2.5-24B-A2B

The largest and final LFM2.5 model, a reasoning-only mixture-of-experts model with 24B total and about 2B active parameters, extending LFM2-24B-A2B with a 131K context window, larger vocabulary, and agentic reinforcement learning.

- Date: 2026-10-13
- Lab: Liquid AI
- Type: Reasoning
- Weights: Open weights
- Modalities: text → text
- Context window: 131,072 tokens
- Parameters: MoE: 24.0B total, 2.45B active
- License: LFM Open License v1.0
- Available at launch: API, Consumer app, Open weights
- Permalink: https://theaiengineering.com/releases/2026/10#lfm2-5-24b-a2b

Highlights:

- Context extended from 32,768 to 131,072 tokens; vocabulary doubled from 65K to 128K
- Pre-training scaled from 17T to 33T tokens
- Agentic RL inside agent runtimes such as OpenClaw, Hermes Agent, Pi, and OpenCode
- Base and post-trained weights on Hugging Face; also on Liquid Playground and Liquid API

Source: [LFM2.5-24B-A2B: Built for Fast On-Device Agents](https://www.liquid.ai/blog/lfm2-5-24b-a2b)

### Muse Wonder and Muse Wonder Full

Muse Wonder and Muse Wonder Full open the Muse Wonder series, Meta's most capable models, aimed at personal and professional agentic tasks with Active Memory and Active Focus features.

- Date: 2026-10-12
- Lab: Meta
- Type: Reasoning
- Weights: Proprietary
- Modalities: text, image, video, audio → text
- Context window: 1,048,576 tokens
- Available at launch: API, Consumer app
- Permalink: https://theaiengineering.com/releases/2026/10#muse-wonder

Highlights:

- Muse Wonder is available through the Meta Model API and Muse Code
- Muse Wonder Full is available to Muse Code subscribers only
- Meta Model API docs list a 1,048,576-token context window for Muse Wonder
- Muse Wonder is slated to power the Muse personal agent, first for subscribers

Source: [Muse Wonder: A Frontier Model For Personal & Professional AI Agents](https://research.meta.ai/blog/muse-wonder)

### Claude Haiku 5.5

Claude Haiku 5.5 is the small, high-volume model in Anthropic's Claude 5.5 family, aimed at summaries, classification, subagent work, and latency-sensitive tasks such as customer support and browser use.

- Date: 2026-10-07
- Lab: Anthropic
- Type: Small model
- Weights: Proprietary
- Modalities: text, image → text
- Context window: 1,000,000 tokens
- Available at launch: API, Consumer app, Cloud platforms
- Permalink: https://theaiengineering.com/releases/2026/10#claude-haiku-5-5

Highlights:

- API model ID claude-haiku-5-5
- 1M-token context window and 128K max output tokens
- First Haiku-class model with an adjustable effort setting
- Available in the Claude apps and on Amazon Web Services, Google Cloud, and Microsoft Azure

Source: [Introducing Claude Haiku 5.5](https://www.anthropic.com/claude-haiku-5-5)

### d1-3B and d1-omni-600M

The first open-weight d1 decision models: d1-3B, trained from LFM2.5-VL-3B for text and image inputs, and the experimental d1-omni-600M, built on LFM2.5-Encoder-350M with added vision and audio encoders.

- Date: 2026-10-07
- Lab: Liquid AI
- Type: Small model
- Weights: Open weights
- Modalities: text, image, audio → score
- Parameters: d1-3B: 3.12B; d1-omni-600M: 587M
- License: LFM Open License v1.0
- Available at launch: Open weights
- Permalink: https://theaiengineering.com/releases/2026/10#d1-3b-omni-600m

Highlights:

- Answers questions in one forward pass without generating tokens
- d1-omni-600M accepts text with images or text with audio
- Context: 32,768 tokens (d1-3B) and 16,384 tokens (d1-omni-600M) per model cards
- Available on Hugging Face with day-one llama.cpp support

Source: [Open d1: Edge decision models for text, vision, and audio](https://www.liquid.ai/blog/d1-open)

### GPT-6 in ChatGPT (GPT-6 Instant, with Intelligent UI)

GPT-6 in ChatGPT brings the GPT-6 generation to the Chat tab for all tiers, introducing Intelligent UI, which composes answers from text, visuals, and interactive elements, and letting ChatGPT begin answering while it keeps thinking.

- Date: 2026-10-07
- Lab: OpenAI
- Type: LLM
- Weights: Proprietary
- Modalities: text, image → text
- Available at launch: Consumer app, Enterprise
- Permalink: https://theaiengineering.com/releases/2026/10#gpt-6-in-chatgpt

Highlights:

- Powered by GPT-6 Sol for Plus, Pro, Business, and Enterprise, and GPT-6 Luna for Free and Go
- Rolled out to paid tiers on October 7, 2026, expanding to Free and Go tiers the next day
- Intelligent UI responses can include graphics, tappable buttons, forms, charts, and interactive tools
- Models powering ChatGPT Work and Codex did not change with this release

Source: [GPT-6 and Intelligent UI for everyone | OpenAI](https://openai.com/index/gpt-6-for-everyone/)

### pplx-embed-v2-late-0.6b and pplx-embed-v2-late-9b

A family of late-interaction (ColBERT-style) embedding models in 0.6B and 9B sizes that output 128-dimensional vectors per token for text and image retrieval, including rendered PDF pages, in a shared embedding space.

- Date: 2026-10-07
- Lab: Perplexity
- Type: Embedding
- Weights: Open weights
- Modalities: text, image → embedding
- Parameters: 0.6b: 594M total; 9b: 7.4B active
- License: MIT
- Available at launch: Open weights
- Permalink: https://theaiengineering.com/releases/2026/10#pplx-embed-v2-late

Highlights:

- Scores query-document pairs with MaxSim over 128-dimensional token embeddings
- The 0.6B model can query an index built with the 9B model through the shared embedding space
- Built on Qwen3.5 with bidirectional attention; the 0.6B model is pruned from Qwen3.5-0.8B
- Weights on Hugging Face under the MIT license

Source: [Multimodal embeddings beyond a single vector](https://www.perplexity.ai/hub/blog/multimodal-embeddings-beyond-a-single-vector)

### EmbeddingGemma 2

EmbeddingGemma 2 is a lightweight open multimodal embedding model built on the Gemma 4 architecture. It maps text, code, images, audio and video into one shared embedding space, succeeding the text-only EmbeddingGemma.

- Date: 2026-10-06
- Lab: Google
- Type: Embedding
- Weights: Open weights
- Modalities: text, image, audio, video → embedding
- Context window: 8,192 tokens
- Parameters: 740M (270M text backbone)
- License: Apache-2.0
- Available at launch: Open weights
- Permalink: https://theaiengineering.com/releases/2026/10#embeddinggemma-2

Highlights:

- 8K-token context window, four times larger than EmbeddingGemma 1
- 768-dimension output with Matryoshka truncation to 512, 256 or 128
- Text-only use needs 270M parameters; vision (170M) and audio (300M) encoders are optional
- Weights on Hugging Face and Kaggle under Apache 2.0

Source: [EmbeddingGemma 2: an open, lightweight multimodal embedding model](https://blog.google/innovation-and-ai/technology/developers-tools/embeddinggemma-2/)

### Mistral Large 4

Mistral's largest model, a natively multimodal mixture-of-experts model nicknamed Le Chonk, launched as an API preview and aimed at agentic coding, cybersecurity and knowledge work.

- Date: 2026-10-06
- Lab: Mistral AI
- Type: Multimodal
- Weights: Open weights
- Modalities: text, image → text
- Context window: 1,000,000 tokens
- Parameters: MoE: 1T total, 52B active
- License: Mistral Ecosystem Open-Weight License
- Available at launch: Open weights, API
- Permalink: https://theaiengineering.com/releases/2026/10#mistral-large-4

Highlights:

- Improved weights released October 29, 2026 under a modified MIT license
- Three reasoning modes balance cost and performance
- Self-hostable with vLLM or SGLang

Source: [Introducing Mistral Large 4 | Mistral](https://mistral.ai/news/mistral-large-4)

### Nano Banana 2.1

Nano Banana 2.1 is an update to Nano Banana 2 that keeps Flash-level speed while improving visual quality, prompt adherence, multi-turn character consistency and text rendering for image generation and editing.

- Date: 2026-10-06
- Lab: Google
- Type: Image
- Weights: Proprietary
- Modalities: text, image, video → text, image
- Context window: 131,072 tokens
- Available at launch: API
- Permalink: https://theaiengineering.com/releases/2026/10#nano-banana-2-1

Highlights:

- Model ID gemini-nano-banana-2.1, released as generally available in the Gemini API
- Adds wide and panoramic aspect ratios (1:4, 4:1, 1:8 and 8:1) at 1K, 2K and 4K
- gemini-3.1-flash-image (Nano Banana 2) was deprecated the same day

Source: [Release notes - Gemini API | Google AI Studio](https://ai.google.dev/gemini-api/docs/changelog)

### Amazon Nova 2.5 Sonic

Amazon Nova 2.5 Sonic is Amazon's speech-to-speech model for real-time voice agents, succeeding Nova 2 Sonic with improved reasoning, instruction following and tool-calling accuracy plus lower response latency.

- Date: 2026-10-05
- Lab: Amazon
- Type: Speech
- Weights: Proprietary
- Modalities: text, audio → text, audio
- Context window: 256,000 tokens
- Available at launch: API
- Permalink: https://theaiengineering.com/releases/2026/10#amazon-nova-2-5-sonic

Highlights:

- Generally available in Amazon Bedrock at launch
- 256K token context window with up to 64K output tokens
- Expressive voices in seven languages, with asynchronous tool calling and mixed voice and text sessions

Source: [Announcing Amazon Nova 2.5 Sonic with improved reasoning for voice agents](https://aws.amazon.com/about-aws/whats-new/2026/10/amazon-nova-2.5-Sonic/)

### d1

Liquid AI's first decision model, which reads text and images with one or more questions in a single forward pass and returns probabilities for each answer instead of generating tokens; this release adds image input.

- Date: 2026-10-05
- Lab: Liquid AI
- Type: Multimodal
- Weights: Proprietary
- Modalities: text, image → score
- Available at launch: API
- Permalink: https://theaiengineering.com/releases/2026/10#d1-decision-model

Highlights:

- Liquid API model ID d1, served through the decisions endpoint
- Question types: Noul (yes/no), Choice, and Score
- Also on OpenRouter and Vercel AI Gateway, text only at launch
- Follows an experimental text-only d1 release the prior week

Source: [Introducing d1: The most capable decision model, now with vision](https://www.liquid.ai/blog/d1-decision-model)

### AstaBrief 8B

Open-weights model that turns a research question and retrieved literature excerpts into a cited scientific report, powering Fast mode in Asta's report generation.

- Date: 2026-10-02
- Lab: Ai2
- Type: LLM
- Weights: Open weights
- Modalities: text → text
- Parameters: 8B
- License: Apache-2.0
- Available at launch: Open weights
- Permalink: https://theaiengineering.com/releases/2026/10#astabrief

Highlights:

- Post-trained from Qwen3-8B
- Training data open-sourced with the model
- Released under Apache 2.0

Source: [Open-sourcing AstaBrief, the fast report-generation model in Asta | Ai2](https://allenai.org/blog/astabrief)

### FLUX 3 Image

Image generation and editing endpoint built on FLUX 3 that uses one prompt for both modes, with bounding boxes for layout control, pixel-exact local edits and up to ten reference images.

- Date: 2026-10-01
- Lab: Black Forest Labs
- Type: Image
- Weights: Proprietary
- Modalities: text, image → image
- Available at launch: API
- Permalink: https://theaiengineering.com/releases/2026/10#flux-3-image

Highlights:

- Output from 768px previews up to 4K
- Bounding boxes are added as a JSON list in the prompt
- Custom output sizes and better single-reference edits added October 22, 2026

Source: [Release Notes - Black Forest Labs](https://docs.bfl.ai/release-notes)

### MAI-Transcribe-2-Streaming

MAI-Transcribe-2-Streaming is Microsoft AI's first streaming speech-to-text model, delivering low-latency, real-time transcripts for conversational voice agents in 60 languages with automatic, continuous language detection.

- Date: 2026-10-01
- Lab: Microsoft
- Type: Speech
- Weights: Proprietary
- Modalities: audio → text
- Available at launch: API, Consumer app, Cloud platforms
- Permalink: https://theaiengineering.com/releases/2026/10#mai-transcribe-2-streaming

Highlights:

- Real-time transcription in 60 languages
- Automatic, continuous language detection
- Available through Microsoft Foundry, MAI Playground, Vercel and Azure Voice Live

Source: [Our first streaming transcription model debuts at no. 1 on Artificial Analysis](https://microsoft.ai/news/our-first-streaming-transcription-model/)

### MAI-Voice-2.1

MAI-Voice-2.1 is Microsoft AI's text-to-speech model announced alongside MAI-Transcribe-2-Streaming for conversational voice agents. Microsoft calls it its strongest multilingual text-to-speech model, now covering 23 languages and 26 locales.

- Date: 2026-10-01
- Lab: Microsoft
- Type: Speech
- Weights: Proprietary
- Modalities: text → audio
- Available at launch: API, Consumer app, Cloud platforms
- Permalink: https://theaiengineering.com/releases/2026/10#mai-voice-2-1

Highlights:

- Supports 23 languages and 26 locales
- Available through Microsoft Foundry, MAI Playground, OpenRouter, Vercel and Azure Voice Live
- Released with a faster MAI-Voice-2.1-Flash variant

Source: [Our first streaming transcription model debuts at no. 1 on Artificial Analysis](https://microsoft.ai/news/our-first-streaming-transcription-model/)

### MAI-Voice-2.1-Flash

MAI-Voice-2.1-Flash is a faster, high-volume variant of Microsoft AI's MAI-Voice-2.1 text-to-speech model, built for latency-sensitive voice agent workloads while supporting the same languages and cross-language speakers.

- Date: 2026-10-01
- Lab: Microsoft
- Type: Speech
- Weights: Proprietary
- Modalities: text → audio
- Available at launch: API, Consumer app, Cloud platforms
- Permalink: https://theaiengineering.com/releases/2026/10#mai-voice-2-1-flash

Highlights:

- Supports the same 23 languages as MAI-Voice-2.1
- Available through Microsoft Foundry, MAI Playground, OpenRouter, Vercel and Azure Voice Live
- Announced alongside MAI-Transcribe-2-Streaming

Source: [Our first streaming transcription model debuts at no. 1 on Artificial Analysis](https://microsoft.ai/news/our-first-streaming-transcription-model/)

### Suno Speech (beta)

Beta audio model in Suno that generates spoken audio set to original background music from text, with the voice and musical style described in the prompt.

- Date: 2026-10-01
- Lab: Suno
- Type: Speech
- Weights: Proprietary
- Modalities: text → audio
- Available at launch: Consumer app
- Permalink: https://theaiengineering.com/releases/2026/10#suno-speech

Highlights:

- Generates voice and music together
- Built into the Suno app
- Released as a beta

Source: [Introducing Speech (beta)](https://suno.com/blog/introducing-speech-beta)

## September 2026

### Embed 5

Cohere's fifth-generation embedding family, in Pro and Fast variants that share an embedding space, for retrieval over visually rich documents, filings, code and multilingual data.

- Date: 2026-09-30
- Lab: Cohere
- Type: Embedding
- Weights: Proprietary
- Modalities: text, image → embedding
- Context window: 128,000 tokens
- Available at launch: API, Cloud platforms, Enterprise
- Permalink: https://theaiengineering.com/releases/2026/09#cohere-embed-5

Highlights:

- Model IDs embed-v5.0-pro and embed-v5.0-fast
- Matryoshka dimensions from 256 to 2048
- Available on Microsoft Foundry and Amazon SageMaker

Source: [Embed 5: Frontier Embedding Models for Enterprise | Cohere](https://cohere.com/blog/embed-5)

### Gemini 4 Argon

Gemini 4 Argon is Google's frontier reasoning model for complex coding, agentic, math, science and enterprise work, with a unified 1 million token input and output limit in the Gemini API.

- Date: 2026-09-30
- Lab: Google
- Type: Reasoning
- Weights: Proprietary
- Modalities: text, image, audio, video → text
- Context window: 1,048,576 tokens
- Available at launch: API, Cloud platforms, Consumer app, Enterprise
- Permalink: https://theaiengineering.com/releases/2026/09#gemini-4-argon

Highlights:

- Initially rolled out to trusted cyber defenders through the Fairwind Program
- Generally available on Oct 13, 2026 in the Gemini API, Google Cloud, Google Antigravity, the Gemini app and AI Mode in Search
- Output limit raised to 1 million tokens from the previous 64K via a long decode continuation protocol

Source: [Introducing Gemini 4 Argon](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/)

### Ideogram 4.5

Ideogram image editing model built for multi-turn editing, changing only what a prompt asks so repeated edits do not build up artifacts, pixel shifts or color changes.

- Date: 2026-09-30
- Lab: Ideogram
- Type: Image
- Weights: Proprietary
- Modalities: text, image → image
- Available at launch: Consumer app, API
- Permalink: https://theaiengineering.com/releases/2026/09#ideogram-4-5

Highlights:

- Four quality modes, all at native 2K resolution
- Edits stylized text in place or translates it
- Ideogram said open weights are coming soon

Source: [Ideogram on X: Introducing Ideogram 4.5, the most precise edit model.](https://x.com/ideogram_ai/status/2105327223431737780)

### pplx-embed-v2-context-9b-preview

A preview contextual embedding model that embeds document chunks with their surrounding context, trained by distilling chunk relevance from Perplexity's context compression model to retrieve supporting context as well as answer chunks.

- Date: 2026-09-30
- Lab: Perplexity
- Type: Embedding
- Weights: Open weights
- Modalities: text → embedding
- Parameters: 9B
- License: MIT
- Available at launch: Open weights
- Permalink: https://theaiengineering.com/releases/2026/09#pplx-embed-v2-context

Highlights:

- Returns one INT8 embedding per chunk, with 1024- or 2048-dimensional (MRL) vectors
- Non-preview pplx-embed-v2-context-0.6b and -9b (text and image) followed on Hugging Face in October 2026
- Weights on Hugging Face under the MIT license

Source: [Contextual embedding beyond the gold passage](https://www.perplexity.ai/hub/blog/contextual-embedding-beyond-the-gold-passage)

### GPT-6.1 Sol

GPT-6.1 Sol is an upgrade to GPT-6 Sol for agentic coding, computer use, and professional work, which OpenAI positions as delivering near GPT-6 Astra intelligence at substantially lower cost.

- Date: 2026-09-29
- Lab: OpenAI
- Type: LLM
- Weights: Proprietary
- Modalities: text, image → text
- Context window: 1,050,000 tokens
- Available at launch: Consumer app, API, Enterprise
- Permalink: https://theaiengineering.com/releases/2026/09#gpt-6-1-sol

Highlights:

- API model ID gpt-6.1-sol in the Responses and Chat Completions APIs; multi-agent delegation in beta
- In ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu users; not in the Chat tab at launch
- 1,050,000-token context window and 128,000 max output tokens per the API model page
- Ultrafast mode for gpt-6.1-sol reached the Responses API on October 8, 2026

Source: [Introducing GPT-6.1 Sol | OpenAI](https://openai.com/index/introducing-gpt-6-1-sol/)

### Claude Sonnet 5.5

Claude Sonnet 5.5 is the second model in Anthropic's Claude 5.5 family, positioned as a faster, lower-cost complement to Claude Opus 5.5 for well-scoped everyday tasks, bug fixing, and documents, slides, and spreadsheets.

- Date: 2026-09-28
- Lab: Anthropic
- Type: LLM
- Weights: Proprietary
- Modalities: text, image → text
- Context window: 1,000,000 tokens
- Available at launch: API, Consumer app, Cloud platforms
- Permalink: https://theaiengineering.com/releases/2026/09#claude-sonnet-5-5

Highlights:

- API model ID claude-sonnet-5-5; 1M-token context window and 128K max output tokens
- First Sonnet model to launch with cyber safeguards and fallbacks
- New between_tools thinking setting replaces disabling thinking
- Default effort is Medium in Claude Code and the Claude apps, High on the Claude Platform

Source: [Introducing Claude Sonnet 5.5](https://www.anthropic.com/claude-sonnet-5-5)

### Eleven v4 and Eleven v4 Turbo

ElevenLabs' most emotive text-to-speech model, built on a new architecture to interpret tone, pacing and character, with a low-latency Eleven v4 Turbo variant for agents.

- Date: 2026-09-28
- Lab: ElevenLabs
- Type: Speech
- Weights: Proprietary
- Modalities: text → audio
- Available at launch: Consumer app, API
- Permalink: https://theaiengineering.com/releases/2026/09#eleven-v4

Highlights:

- Supports 90+ languages
- Eleven v4 Turbo has a median inference latency of about 100ms
- Served through the Text to Dialogue API and WebSocket

Source: [Introducing Eleven v4, our most emotive model](https://elevenlabs.io/blog/eleven-v4)

### LongCat-2.5

LongCat-2.5 adds image understanding to Meituan's LongCat line and targets agentic coding, combining visual understanding with long-horizon engineering tasks. A preview launched on the LongCat API in September, followed by the full release in November 2026.

- Date: 2026-09-25
- Lab: Meituan
- Type: Multimodal
- Weights: Open weights
- Modalities: text, image → text
- Parameters: MoE: 1.6T total, about 48B active
- License: MIT
- Available at launch: API
- Permalink: https://theaiengineering.com/releases/2026/09#longcat-2-5

Highlights:

- LongCat-2.5-Preview became available on the LongCat API on September 25, 2026
- LongCat-2.5 became available on November 2, 2026
- Supports visual coding, creative design, and 3D tasks
- Works with Claude Code, Codex, Hermes, and other coding agents
- Weights published on Hugging Face under the MIT license in November 2026

Source: [Change Log | LongCat API Docs](https://longcat.chat/platform/docs/change-log)

### FLUX 3 Action

Open-weight 7B world action model derived from the FLUX 3 multimodal backbone that predicts robot actions jointly with video, released with weights, code and a report on its fine-tuning recipe.

- Date: 2026-09-23
- Lab: Black Forest Labs
- Type: Multimodal
- Weights: Open weights
- Modalities: text, image, video → video, action
- Parameters: 7B
- License: FLUX Kommunity License
- Available at launch: Open weights
- Permalink: https://theaiengineering.com/releases/2026/09#flux-3-action

Highlights:

- 7B parameters, pretrained on video, image and audio data
- Weights released under the FLUX Kommunity License
- Targets robot control, with gaming explored as a testbed

Source: [FLUX 3 Action: A 7B World Action Model for Robot Control](https://bfl.ai/blog/flux-3-action)

### Gemini 3.8 Flash TTS

Gemini 3.8 Flash TTS is Google's flagship creative text-to-speech model, built for studio-grade voice fidelity, nuanced acting and long-form multi-turn stability, with voices designed from natural language prompts.

- Date: 2026-09-23
- Lab: Google
- Type: Speech
- Weights: Proprietary
- Modalities: text → audio
- Context window: 8,192 tokens
- Available at launch: API, Enterprise
- Permalink: https://theaiengineering.com/releases/2026/09#gemini-3-8-flash-tts

Highlights:

- Voice design across more than 100 languages and dialects, plus voice replication from a 30-second sample with consent verification
- Launched with the Gemini API Voices endpoint and an extended voice library
- Generated audio carries SynthID watermarking

Source: [Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-8-text-to-speech/)

### Gemini 3.8 Flash-Lite TTS

Gemini 3.8 Flash-Lite TTS is a fast, cost-efficient text-to-speech model for high-throughput production use and real-time voice agent pipelines, replacing the Gemini 3.1 Flash TTS preview.

- Date: 2026-09-23
- Lab: Google
- Type: Speech
- Weights: Proprietary
- Modalities: text → audio
- Context window: 8,192 tokens
- Available at launch: API, Enterprise
- Permalink: https://theaiengineering.com/releases/2026/09#gemini-3-8-flash-lite-tts

Highlights:

- Model ID gemini-3.8-flash-lite-tts, generally available in the Gemini API from Sep 22, 2026
- Released alongside Gemini 3.8 Flash TTS and the Gemini API Voices endpoint
- Input token limit of 8,192 per the Gemini API model page

Source: [Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-8-text-to-speech/)

### Recraft V4.1 Flash

Fast variant of Recraft V4.1 that generates an image in about 1.3 seconds from prompt while keeping the Recraft look across photography, portraits, complex scenes and early concepts.

- Date: 2026-09-23
- Lab: Recraft
- Type: Image
- Weights: Proprietary
- Modalities: text → image
- Available at launch: Consumer app, API
- Permalink: https://theaiengineering.com/releases/2026/09#recraft-v4-1-flash

Highlights:

- About 1.3 seconds from prompt to image
- Available in Recraft Studio

Source: [Meet Recraft V4.1 Flash: The Fastest Image Model on the Market](https://www.recraft.ai/blog/meet-recraft-v4-1-flash)

### Claude Opus 5.5

Claude Opus 5.5 is the first model in Anthropic's Claude 5.5 family, built for long-running agentic coding and knowledge work. It ships with safeguards similar to Claude Fable 5.1 and requires less compute to serve than Opus 5.

- Date: 2026-09-22
- Lab: Anthropic
- Type: LLM
- Weights: Proprietary
- Modalities: text, image → text
- Context window: 1,000,000 tokens
- Available at launch: API, Consumer app, Cloud platforms
- Permalink: https://theaiengineering.com/releases/2026/09#claude-opus-5-5

Highlights:

- API model ID claude-opus-5-5; 1M-token context window and 128K max output tokens
- Adaptive thinking is always on and cannot be disabled
- Preserved thinking applies to API accounts created on or after August 31, 2026
- Available on Amazon Web Services, Google Cloud, and Microsoft Azure; zero data retention available

Source: [Introducing Claude Opus 5.5](https://www.anthropic.com/claude-opus-5-5)

### GPT-6 Sol and GPT-6 Luna

GPT-6 Sol and GPT-6 Luna extend the GPT-6 family with lower-cost tiers trained with methods similar to GPT-6 Astra, with Sol aimed at complex coding and agentic workflows and Luna at focused, high-volume tasks.

- Date: 2026-09-22
- Lab: OpenAI
- Type: LLM
- Weights: Proprietary
- Modalities: text, image → text
- Context window: 1,050,000 tokens
- Available at launch: Consumer app, API, Enterprise
- Permalink: https://theaiengineering.com/releases/2026/09#gpt-6-sol

Highlights:

- API model IDs gpt-6-sol and gpt-6-luna, with text and image input and text output
- In ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu; Free and Go get Luna in the desktop app
- Both list a 1,050,000-token context window on their API model pages
- Not available in the ChatGPT Chat tab at launch

Source: [Introducing GPT-6 Sol and Luna | OpenAI](https://openai.com/index/introducing-gpt-6-sol-and-luna/)

### Hy Image 3.5 preview

Hy Image 3.5 preview is Tencent's image generation model building on the Hy Image 3.0 series, with improved text rendering and layout, realism and style range, and editing consistency for everyday and professional visual work.

- Date: 2026-09-22
- Lab: Tencent
- Type: Image
- Weights: Proprietary
- Modalities: text, image → image
- Available at launch: API, Consumer app
- Permalink: https://theaiengineering.com/releases/2026/09#hy-image-3-5-preview

Highlights:

- Supports text-to-image and image-to-image with up to five reference images
- Multi-turn creation and continuous editing, with output up to 2K resolution
- Available through the Tencent Cloud API, Hy AI Studio, Yuanbao, and ima

Source: [Hy Image3.5 preview 发布：为专业创作提供高性价比模型](https://hy.tencent.com/research/hy-image35-preview)

### MiMo-V2.6-Pro and MiMo-V2.6-Flash

MiMo-V2.6 is Xiaomi's open-weight series of two natively omnimodal models, MiMo-V2.6-Pro and MiMo-V2.6-Flash, post-trained with substantially scaled reinforcement learning on verifiable coding, agent, visual, and cyber tasks.

- Date: 2026-09-22
- Lab: Xiaomi
- Type: Multimodal
- Weights: Open weights
- Modalities: text, image, video, audio → text
- Context window: 1,000,000 tokens
- Parameters: MoE: 1.02T total, 42B active (Pro); 309B total, 15B active (Flash)
- License: MIT
- Available at launch: API, Open weights
- Permalink: https://theaiengineering.com/releases/2026/09#mimo-v2-6

Highlights:

- 1M-token context with text, image, video, and audio input
- Weights released under MIT along with the technical report, training environments, and RL code
- Launched with MiMo-V2.6-Pro-UltraSpeed, a faster serving option on the API
- Both models completed a live-streamed RL run of under six days

Source: [MiMo-V2.6 | Xiaomi](https://mimo.xiaomi.com/mimo-v2-6/article.html)

### Scribe v2 Medical

Speech-to-text model fine-tuned from Scribe v2 for clinical audio, handling drug names, dosages and medical terminology, with the same features, languages and API as Scribe v2.

- Date: 2026-09-22
- Lab: ElevenLabs
- Type: Speech
- Weights: Proprietary
- Modalities: audio → text
- Available at launch: API
- Permalink: https://theaiengineering.com/releases/2026/09#elevenlabs-scribe-v2-medical

Highlights:

- Model ID scribe_v2_medical
- Generally available on ElevenAPI
- Built for protected health information workflows

Source: [Scribe v2 Medical is generally available](https://elevenlabs.io/blog/scribe-v2-medical-generally-available)

### Grok 4.7

Grok 4.7 is SpaceXSI's frontier model for coding, agentic tasks, and knowledge work. It uses a new, larger base model than Grok 4.6 and a longer reinforcement learning run weighted toward multi-hour tasks.

- Date: 2026-09-21
- Lab: xAI
- Type: LLM
- Weights: Proprietary
- Modalities: text, image → text
- Context window: 500,000 tokens
- Available at launch: API, Cloud platforms
- Permalink: https://theaiengineering.com/releases/2026/09#grok-4-7

Highlights:

- API model ID: grok-4.7, with a 500k-token context window
- Reasoning effort settings: low, medium, high (default), and xhigh
- Available at launch in Cursor and Grok Build; a fast variant doubles output speed
- Trained to work natively inside Grok Bot

Source: [Introducing Grok 4.7 | SpaceXSI](https://x.ai/news/grok-4-7)

### Qwen-Image-2.1

Qwen-Image-2.1 is an open-weights Qwen image model that unifies text-to-image generation and image editing in a single model. It is designed to balance generation quality, inference efficiency, and cost, and natively supports transparent images.

- Date: 2026-09-20
- Lab: Alibaba Qwen
- Type: Image
- Weights: Open weights
- Modalities: text, image → image
- Parameters: 7B (visual generation component)
- License: Qwen Research License Agreement
- Available at launch: API, Open weights
- Permalink: https://theaiengineering.com/releases/2026/09#qwen-image-2-1

Highlights:

- Visual generation component has 32 single-stream DiT layers and 7B parameters
- Qwen-Image-2.1-Turbo, which needs only 8 denoising steps, followed on October 9, 2026
- Pro and Turbo APIs went live on Alibaba Cloud Model Studio on October 9, 2026

Source: [Qwen-Image-2.1: Compact, Efficient, and Unified Image Creation](https://qwen.ai/blog?id=qwen-image-2.1)

### Step 5 Preview

Step 5 Preview is StepFun's flagship model for agentic work, aimed at software engineering and professional knowledge work with particular strength in finance. It launched through StepFun products and API, with open weights following on October 15, 2026.

- Date: 2026-09-20
- Lab: StepFun
- Type: LLM
- Weights: Open weights
- Modalities: text, image, video → text
- Context window: 1,000,000 tokens
- Parameters: MoE: 600B total, 27B active
- License: StepFun Model License 1.0
- Available at launch: API, Consumer app, Open weights
- Permalink: https://theaiengineering.com/releases/2026/09#step-5-preview

Highlights:

- Sparse mixture-of-experts model with 600B total and 27B active parameters
- 1M-token context window; accepts text, image, and video input
- API model ID step-5-preview with low, medium, and high reasoning effort
- BF16 and FP8 weights on Hugging Face under the StepFun Model License 1.0

Source: [阶跃发布旗舰模型 Step 5 Preview：向前一步，智能效率的新一代“帕累托前沿”](https://www.stepfun.com/news/step-5-preview)

### Grok Voice Transcribe 2.0

Grok Voice Transcribe 2.0 is SpaceXSI's speech-to-text model, built on the audio foundation model behind Grok Voice. It replaces Grok Voice Transcribe 1.0 as the default Speech to Text model, with its largest gains in multilingual accuracy.

- Date: 2026-09-18
- Lab: xAI
- Type: Speech
- Weights: Proprietary
- Modalities: audio → text
- Available at launch: API
- Permalink: https://theaiengineering.com/releases/2026/09#grok-voice-transcribe-2-0

Highlights:

- API model ID: grok-voice-transcribe-2.0, the default Speech to Text model
- Batch and streaming modes, word-level timestamps, and speaker diarization
- Detects language automatically and follows mid-recording language switches
- grok-voice-transcribe-1.0 requests route to 2.0 from October 2, 2026

Source: [Introducing Grok Voice Transcribe 2.0 | SpaceXSI](https://x.ai/news/grok-voice-transcribe-2)

### Qwen3.8-Omni-Flash

Qwen3.8-Omni-Flash is Qwen's omni-modal model for real-time audio-visual interaction and agentic delivery. It accepts text, image, audio, and video input with a 1M-token context window.

- Date: 2026-09-18
- Lab: Alibaba Qwen
- Type: Multimodal
- Weights: Proprietary
- Modalities: text, image, audio, video → text, audio
- Context window: 1,000,000 tokens
- Available at launch: API
- Permalink: https://theaiengineering.com/releases/2026/09#qwen3-8-omni-flash

Highlights:

- Qwen reports text performance comparable to a text-only model of the same size
- Upgraded video captioning lets users set subject, time range, detail, and format

Source: [Qwen3.8-Omni-Flash: Omni Senses. Agentic Delivery.](https://qwen.ai/blog?id=qwen3.8-omni-flash)

### LFM2-1.2B-Longevity and LFM2-2.6B-Longevity

Domain-adapted LFM2 models fine-tuned on aging-biology data spanning clinical records, DNA methylation, transcriptomics, proteomics, and genetics, released with the LongevityBench benchmark and an accompanying Cell paper.

- Date: 2026-09-17
- Lab: Liquid AI
- Type: Small model
- Weights: Open weights
- Modalities: text → text
- Context window: 32,768 tokens
- Parameters: 1.2B; 2.6B
- License: LFM Open License v1.0
- Available at launch: Open weights
- Permalink: https://theaiengineering.com/releases/2026/09#lfm2-longevity

Highlights:

- Fine-tuned from LFM2-1.2B and LFM2-2.6B for three epochs at a 32k-token context
- LongevityBench covers 17 tasks and 25,457 prompts across five biodata domains
- Weights available on Hugging Face

Source: [Setting the Frontier of Aging Biology with Liquid Foundation Models](https://www.liquid.ai/blog/longevitybench)

### Gemini 3.8 Live

Gemini 3.8 Live is a native audio model for real-time conversation in the Live API, built for sub-second latency, near real-time visual context and background tool use without interrupting the conversation.

- Date: 2026-09-15
- Lab: Google
- Type: Speech
- Weights: Proprietary
- Modalities: text, image, audio → text, audio
- Context window: 131,072 tokens
- Available at launch: API, Consumer app, Enterprise
- Permalink: https://theaiengineering.com/releases/2026/09#gemini-3-8-live

Highlights:

- Detects and switches between 97 supported languages mid-conversation
- Available in the Gemini API, Google Workspace and the Gemini app
- Gemini Enterprise added a Live Avatar feature on Sep 24, 2026

Source: [Gemini 3.8 Live & Gemini 3.8 Live Extended Thinking](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-8-live-gemini-3-8-live-extended-thinking/)

### Gemini 3.8 Live Extended Thinking

Gemini 3.8 Live Extended Thinking is a speech-to-speech Live API model specialized for complex reasoning, multi-step planning and long-horizon state tracking during real-time voice conversations.

- Date: 2026-09-15
- Lab: Google
- Type: Speech
- Weights: Proprietary
- Modalities: text, image, audio → text, audio
- Context window: 131,072 tokens
- Available at launch: API
- Permalink: https://theaiengineering.com/releases/2026/09#gemini-3-8-live-extended-thinking

Highlights:

- Model ID gemini-3.8-live-extended-thinking with configurable thinking levels from MINIMAL to HIGH
- Supports asynchronous function calling
- Input token limit of 131,072 per the Gemini API model page

Source: [Gemini 3.8 Live & Gemini 3.8 Live Extended Thinking](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-8-live-gemini-3-8-live-extended-thinking/)

### Jev

TypeSafe AI's first public model and the first of what it calls System One models: it reads unstructured data and program state and returns type-safe structured values with calibrated probabilities instead of generating text.

- Date: 2026-09-15
- Lab: TypeSafe AI
- Type: LLM
- Weights: Proprietary
- Modalities: text → score
- Available at launch: API
- Permalink: https://theaiengineering.com/releases/2026/09#jev

Highlights:

- No string generation: answers come back as typed values with confidence scores
- Trained with what TypeSafe calls Reinforcement Learning for Calibrated Decisions (RLCD)
- TypeSafe reported 70–500 ms end-to-end response times on its own benchmarks
- Launched in early access on TypeSafe's hosted API at $0.042 per million input tokens, with output tokens free

Source: [Introducing System One Models & Jev](https://typesafe.ai/blog/introducing-system-one-models-and-jev)

### StepAudio 3

StepAudio 3 is StepFun's speech and audio model family, comprising StepAudio 3 Realtime, ASR, TTS, Gen, and Music for full-duplex voice interaction, speech recognition, speech synthesis, multi-element audio generation, and music creation.

- Date: 2026-09-15
- Lab: StepFun
- Type: Speech
- Weights: Proprietary
- Modalities: text, audio → text, audio
- Available at launch: API
- Permalink: https://theaiengineering.com/releases/2026/09#stepaudio-3

Highlights:

- StepAudio 3 Realtime supports native full-duplex conversation with asynchronous tool calls
- StepAudio 3 Gen produces voices, sound effects, ambience, and background music from text descriptions and reference audio
- StepAudio 3 Music supports song creation, a cappella accompaniment, and covers, with ABC notation control
- All five models launched on the StepFun open platform

Source: [阶跃发布语音大模型StepAudio 3，拿下多个全球第一！](https://www.stepfun.com/news/stepaudio-3)

### Music v2.5

Most advanced ElevenLabs music model, producing more layered and complex songs with more natural-sounding instruments, and the default for prompted and reference generation in ElevenMusic.

- Date: 2026-09-11
- Lab: ElevenLabs
- Type: Music
- Weights: Proprietary
- Modalities: text, audio → audio
- Available at launch: Consumer app, API
- Permalink: https://theaiengineering.com/releases/2026/09#eleven-music-v2-5

Highlights:

- Model ID music_v2_5
- Default model in ElevenMusic
- Improved quality and prompt adherence over Music v2

Source: [Music v2.5](https://elevenlabs.io/blog/music-v2-5-model)

### DeepSeek-V4.1-Flash

DeepSeek-V4.1-Flash is the smallest model in DeepSeek's new architecture family, a multimodal mixture-of-experts model with native image understanding. It uses a Causal Encoder-Decoder design that activates fewer parameters for input than for output.

- Date: 2026-09-10
- Lab: DeepSeek
- Type: Multimodal
- Weights: Open weights
- Modalities: text, image → text
- Context window: 1,000,000 tokens
- Parameters: MoE: 552B backbone, 8B active (prefill), 16B active (decode)
- License: MIT
- Available at launch: API, Open weights
- Permalink: https://theaiengineering.com/releases/2026/09#deepseek-v4-1-flash

Highlights:

- Pre-trained from scratch on a 45T-token multimodal corpus
- Global KV cache about 4 times smaller than DeepSeek-V4-Flash
- Replaced V4-Flash and V4-Flash-Vision-Exp; API model name deepseek-flash
- Supports a continuously controllable reasoning effort from 1 to 100

Source: [DeepSeek-V4.1-Flash: Smarter, Faster, More Efficient](https://api-docs.deepseek.com/news/news260910)

### North Small Translate

Open-weights mixture-of-experts model from Cohere built specifically for machine translation across more than 50 languages and locale variants, with weights offered for non-commercial use.

- Date: 2026-09-10
- Lab: Cohere
- Type: LLM
- Weights: Open weights
- Modalities: text → text
- Context window: 16,000 tokens
- Parameters: MoE: 218B total, 25B active
- License: CC-BY-NC-4.0
- Available at launch: API, Open weights
- Permalink: https://theaiengineering.com/releases/2026/09#north-small-translate

Highlights:

- Available through the Chat V2 API
- Weights in W4A16, FP8 and BF16 for non-commercial use
- Runs on two H100s or one B200 in W4A16

Source: [Introducing North Small Translate | Cohere](https://cohere.com/blog/north-small-translate)

### AuK

AuK is an open 1.5B foundation model from Tencent Hunyuan for speech generation and editing, trained on millions of hours of audio. It supports zero-shot and instruction-based TTS, content and acoustic editing, and paralinguistic editing.

- Date: 2026-09-09
- Lab: Tencent
- Type: Speech
- Weights: Open weights
- Modalities: text, audio → audio
- Parameters: 1.5B
- License: MIT
- Available at launch: Open weights
- Permalink: https://theaiengineering.com/releases/2026/09#hunyuan-auk

Highlights:

- Released together with AuK-Flash, a distilled variant
- Weights on Hugging Face, with demo spaces on Hugging Face and ModelScope
- Day-zero serving support in SGLang-Omni

Source: [GitHub - Tencent-Hunyuan/AuK: AuK: An Open-Source Foundational Model for Speech Generation and Editing](https://github.com/Tencent-Hunyuan/AuK)

### Suno v6 (v6, v6-wild, v6-mini)

New generation of Suno music models developed with industry partners including Warner Music Group, BMG and Believe, offered as flagship v6, exploratory v6-wild and fast, free v6-mini.

- Date: 2026-09-09
- Lab: Suno
- Type: Music
- Weights: Proprietary
- Modalities: text, audio, image, video → audio
- Available at launch: Consumer app
- Permalink: https://theaiengineering.com/releases/2026/09#suno-v6

Highlights:

- Up to 8 minutes of music in a single generation
- v6 and v6-wild for Pro and Premier; v6-mini for all plans
- Previous models to be retired as v6 rolls out

Source: [Introducing v6](https://suno.com/blog/introducing-v6)

### GPT Image 2.5 Sunburst and Flare (ChatGPT Images 2.5)

ChatGPT Images 2.5 is OpenAI's updated image model with sharper detail, more precise multi-turn editing, and lower latency, offered in the API as GPT Image 2.5 Flare for everyday generation and Sunburst for precision editing.

- Date: 2026-09-08
- Lab: OpenAI
- Type: Image
- Weights: Proprietary
- Modalities: text, image → image
- Available at launch: Consumer app, API
- Permalink: https://theaiengineering.com/releases/2026/09#gpt-image-2-5

Highlights:

- API model IDs gpt-image-2.5-flare and gpt-image-2.5-sunburst in the Image API and Responses image tool
- Available to ChatGPT, ChatGPT Work, and Codex users across all tiers on desktop, mobile, and web
- OpenAI reports up to 50% lower image generation latency than Images 2.0
- Both API models support new xhigh and max quality settings

Source: [Introducing ChatGPT Images 2.5 | OpenAI](https://openai.com/index/introducing-chatgpt-images-2-5/)

### MAI-Image-2.6-Flash

MAI-Image-2.6-Flash is a faster variant of Microsoft AI's MAI-Image-2.6, built for latency-sensitive, high-throughput image generation and editing workloads. It launched in Microsoft Foundry alongside developer access to MAI-Image-2.6.

- Date: 2026-09-04
- Lab: Microsoft
- Type: Image
- Weights: Proprietary
- Modalities: text, image → image
- Available at launch: API
- Permalink: https://theaiengineering.com/releases/2026/09#mai-image-2-6-flash

Highlights:

- Supports multi-image reference editing, web grounding and dynamic aspect ratios
- Available to developers in Microsoft Foundry
- Generally available in Foundry per an October 27, 2026 update

Source: [Pushing the quality-cost frontier with MAI-Image-2.6](https://microsoft.ai/news/pushing-the-quality-cost-frontier-with-mai-image-2-6/)

### GPT-6 Astra

GPT-6 Astra is the first model of OpenAI's GPT-6 generation, positioned for the hardest end-to-end work across reasoning, coding, computer use, research, and document creation, with added misalignment monitoring during agent tasks.

- Date: 2026-09-03
- Lab: OpenAI
- Type: LLM
- Weights: Proprietary
- Modalities: text, image → text
- Context window: 1,050,000 tokens
- Available at launch: Consumer app, API, Cloud platforms, Enterprise
- Permalink: https://theaiengineering.com/releases/2026/09#gpt-6-astra

Highlights:

- API model ID gpt-6-astra; tool calling requires the Responses API and the none reasoning effort is unsupported
- Rolled out to a limited set of organizations, then ChatGPT Plus, Pro, Business, Enterprise, Azure, and Bedrock
- GPT-6 Astra Pro offered to Pro, Business, and Enterprise users
- 1,050,000-token context window and 128,000 max output tokens per the API model page

Source: [GPT-6 Astra: A new generation of intelligence | OpenAI](https://openai.com/index/gpt-6-astra/)

### K2 Horizon

K2 Horizon is a family of six fully open foundation models from 0.9B to 375B parameters, released by MBZUAI's Institute of Foundation Models with weights, code, training data, and methodology for reasoning, math, coding, and agentic tasks.

- Date: 2026-09-03
- Lab: MBZUAI IFM
- Type: LLM
- Weights: Open weights
- Parameters: 0.9B to 375B; largest is MoE with 375B total, 23B active
- License: Apache-2.0
- Available at launch: Open weights, API
- Permalink: https://theaiengineering.com/releases/2026/09#k2-horizon

Highlights:

- Sizes: 0.9B, 3.7B, 7B, and 32B dense models, a 36B-A4B sparse model using Mixture of Value Attention (MoVA), and a 375B-A23B flagship
- A diffusion distillation technique generates blocks of tokens in parallel, which IFM says speeds up generation by roughly 3x
- The models share one architecture, vocabulary, and training method, with a smaller vocabulary for the 0.9B model
- Weights on Hugging Face with vLLM and SGLang support, and an API through inference partners including Compass, Cerebras, and Nebius
- Models and code released under Apache 2.0

Source: [K2 Horizon Press Release | Institute of Foundation Models](https://ifm.ai/k2/press-release/)

### MAI-Transcribe-2

MAI-Transcribe-2 is Microsoft AI's batch speech-to-text model, adding speaker diarization, word-level timestamps and configurable transcription styles, with accuracy across 60 languages, up from 43 in MAI-Transcribe-1.5.

- Date: 2026-09-03
- Lab: Microsoft
- Type: Speech
- Weights: Proprietary
- Modalities: audio → text
- Available at launch: API, Consumer app, Cloud platforms
- Permalink: https://theaiengineering.com/releases/2026/09#mai-transcribe-2

Highlights:

- Covers 60 languages
- Adds speaker diarization, word-level timestamps, keyword biasing and automatic language identification
- Available through Microsoft Foundry, MAI Playground and OpenRouter

Source: [MAI-Transcribe-2 is the fastest, most accurate and cheapest speech recognition model in the world](https://microsoft.ai/news/mai-transcribe-2-is-the-fastest-most-accurate-and-cheapest-speech-recognition-model-in-the-world/)

### Gemini 3.8 Flash

Gemini 3.8 Flash is Google's most intelligent Flash model at release, with significant gains over 3.7 Flash in long-horizon software engineering, autonomous agents and multi-step reasoning in specialized domains.

- Date: 2026-09-02
- Lab: Google
- Type: LLM
- Weights: Proprietary
- Modalities: text, image, audio, video → text
- Context window: 1,048,576 tokens
- Available at launch: API, Cloud platforms, Consumer app
- Permalink: https://theaiengineering.com/releases/2026/09#gemini-3-8-flash

Highlights:

- Generally available in the Gemini API via Google AI Studio and Android Studio, in Google Antigravity and in Stitch
- Applies extra reasoning steps and iterative tool calls on complex tasks, with lower effort levels to limit token use
- Announced with Gemini 3.8 Flash Cyber, which is limited to trusted defenders through the Fairwind Program

Source: [Introducing Gemini 3.8 Flash and 3.8 Flash Cyber](https://blog.google/innovation-and-ai/models-and-research/gemini-models/3-8-flash-and-3-8-flash-cyber/)

### Muse Spark 1.3

Muse Spark 1.3 is an update to Meta's Muse Spark reasoning model that improves long-horizon agentic and coding work, follows long instructions more reliably, and adds a max reasoning setting.

- Date: 2026-09-02
- Lab: Meta
- Type: Reasoning
- Weights: Proprietary
- Modalities: text, image, video, audio → text
- Context window: 1,048,576 tokens
- Available at launch: API
- Permalink: https://theaiengineering.com/releases/2026/09#muse-spark-1-3

Highlights:

- Available with max reasoning on Muse Code and the Meta Model API
- Takes fewer unnecessary turns and is less verbose in coding than Muse Spark 1.2
- Meta reports stronger adversarial robustness, including against prompt injection
- API docs list a 1,048,576-token context window; audio input is not fully supported

Source: [Introducing Muse Spark 1.3](https://research.meta.ai/blog/introducing-muse-spark-1-3)

### Claude Fable 5.1

Claude Fable 5.1 is the successor to Claude Fable 5 for long-running agentic coding, knowledge work, and research. It is the generally available version of the model behind Claude Mythos 5.1, with safeguards for cybersecurity and biology.

- Date: 2026-09-01
- Lab: Anthropic
- Type: LLM
- Weights: Proprietary
- Modalities: text, image → text
- Context window: 1,000,000 tokens
- Available at launch: API, Consumer app, Cloud platforms
- Permalink: https://theaiengineering.com/releases/2026/09#claude-fable-5-1

Highlights:

- API model ID claude-fable-5-1; 1M-token context window and 128K max output tokens
- Claude Mythos 5.1, the same model with fewer safeguards, is limited to vetted organizations
- tool_choice types any and tool are not supported
- Available on all platforms, including Amazon Web Services, Google Cloud, and Microsoft Azure

Source: [Introducing Claude Fable 5.1 and Claude Mythos 5.1](https://www.anthropic.com/claude-fable-and-mythos-5-1)

### Muse Voice Transcribe

Muse Voice Transcribe is Meta Superintelligence Labs' first real-time audio perception model, combining streaming speech recognition, speaker diarization and endpointing in one model.

- Date: 2026-09-01
- Lab: Meta
- Type: Speech
- Weights: Proprietary
- Modalities: audio → text
- Available at launch: API, Consumer app
- Permalink: https://theaiengineering.com/releases/2026/09#muse-voice-transcribe

Highlights:

- Diarization for more than 20 speakers
- Trained on 70+ languages, with 25 extensively verified, and handles code-switching
- Available in the Meta Model API, Meta AI for Mac and Muse Code
- API model ID muse-voice-transcribe-1.0

Source: [Introducing Muse Voice Transcribe](https://research.meta.ai/blog/introducing-muse-voice-transcribe)

### PII-Tracer

A compact 0.6B model for detecting personally identifiable information in long, multilingual conversations, released with the PII-TRACE benchmark and built for the local privacy gate in Perplexity Computer's hybrid compute on Mac.

- Date: 2026-09-01
- Lab: Perplexity
- Type: Small model
- Weights: Open weights
- Modalities: text → text
- Parameters: 0.6B
- License: MIT
- Available at launch: Open weights
- Permalink: https://theaiengineering.com/releases/2026/09#pii-tracer

Highlights:

- Bidirectional Qwen3 encoder initialized from pplx-embed-v1-0.6b, with token-tagging and sensitivity heads
- Tags 9 PII categories, including names, account numbers, addresses, emails and secrets
- Released with PII-TRACE, a 13-language benchmark for PII detection across conversations
- Weights on Hugging Face under the MIT license, plus GGUF and MLX builds

Source: [PII-TRACE: Detecting Personal Data Before It Leaves a Device](https://www.perplexity.ai/hub/blog/pii-trace-detecting-personal-data-before-it-leaves-the-device)

## Browse by lab

- [Google](https://theaiengineering.com/releases/google): 106 releases
- [Alibaba Qwen](https://theaiengineering.com/releases/alibaba-qwen): 84 releases
- [OpenAI](https://theaiengineering.com/releases/openai): 79 releases
- [Mistral AI](https://theaiengineering.com/releases/mistral): 59 releases
- [Meta](https://theaiengineering.com/releases/meta): 44 releases
- [Z.ai](https://theaiengineering.com/releases/z-ai): 43 releases
- [Microsoft](https://theaiengineering.com/releases/microsoft): 42 releases
- [Tencent](https://theaiengineering.com/releases/tencent): 42 releases
- [ByteDance Seed](https://theaiengineering.com/releases/bytedance-seed): 40 releases
- [DeepSeek](https://theaiengineering.com/releases/deepseek): 38 releases
- [Liquid AI](https://theaiengineering.com/releases/liquid-ai): 34 releases
- [Cohere](https://theaiengineering.com/releases/cohere): 31 releases
- [IBM](https://theaiengineering.com/releases/ibm): 30 releases
- [xAI](https://theaiengineering.com/releases/xai): 30 releases
- [NVIDIA](https://theaiengineering.com/releases/nvidia): 29 releases
- [MiniMax](https://theaiengineering.com/releases/minimax): 28 releases
- [Stability AI](https://theaiengineering.com/releases/stability-ai): 28 releases
- [Anthropic](https://theaiengineering.com/releases/anthropic): 26 releases
- [Ai2](https://theaiengineering.com/releases/ai2): 23 releases
- [StepFun](https://theaiengineering.com/releases/stepfun): 22 releases
- [Amazon](https://theaiengineering.com/releases/amazon): 21 releases
- [Baidu](https://theaiengineering.com/releases/baidu): 21 releases
- [ElevenLabs](https://theaiengineering.com/releases/elevenlabs): 21 releases
- [Black Forest Labs](https://theaiengineering.com/releases/black-forest-labs): 18 releases
- [Kuaishou](https://theaiengineering.com/releases/kuaishou): 16 releases
- [Meituan](https://theaiengineering.com/releases/meituan): 16 releases
- [Runway](https://theaiengineering.com/releases/runway): 16 releases
- [Moonshot AI](https://theaiengineering.com/releases/moonshot-ai): 15 releases
- [Midjourney](https://theaiengineering.com/releases/midjourney): 14 releases
- [Salesforce](https://theaiengineering.com/releases/salesforce): 13 releases
- [Xiaomi](https://theaiengineering.com/releases/xiaomi): 13 releases
- [Nous Research](https://theaiengineering.com/releases/nous-research): 11 releases
- [Perplexity](https://theaiengineering.com/releases/perplexity): 10 releases
- [AI21 Labs](https://theaiengineering.com/releases/ai21): 8 releases
- [Apple](https://theaiengineering.com/releases/apple): 8 releases
- [Ideogram](https://theaiengineering.com/releases/ideogram): 8 releases
- [01.AI](https://theaiengineering.com/releases/01-ai): 7 releases
- [Hugging Face](https://theaiengineering.com/releases/hugging-face): 7 releases
- [Lightricks](https://theaiengineering.com/releases/lightricks): 7 releases
- [Luma AI](https://theaiengineering.com/releases/luma-ai): 7 releases
- [Reka](https://theaiengineering.com/releases/reka): 7 releases
- [Suno](https://theaiengineering.com/releases/suno): 7 releases
- [Databricks](https://theaiengineering.com/releases/databricks): 5 releases
- [Leonardo.Ai](https://theaiengineering.com/releases/leonardo-ai): 5 releases
- [Pika](https://theaiengineering.com/releases/pika): 5 releases
- [Recraft](https://theaiengineering.com/releases/recraft): 5 releases
- [Inflection AI](https://theaiengineering.com/releases/inflection-ai): 3 releases
- [MBZUAI IFM](https://theaiengineering.com/releases/mbzuai-ifm): 3 releases
- [Snowflake](https://theaiengineering.com/releases/snowflake): 3 releases
- [Thinking Machines Lab](https://theaiengineering.com/releases/thinking-machines-lab): 3 releases
- [TypeSafe AI](https://theaiengineering.com/releases/typesafe-ai): 2 releases
- [Cognition](https://theaiengineering.com/releases/cognition): 1 releases
- [Genmo](https://theaiengineering.com/releases/genmo): 1 releases
- [Reflection AI](https://theaiengineering.com/releases/reflection-ai): 1 releases
- [Ricursive Intelligence](https://theaiengineering.com/releases/ricursive-intelligence): 1 releases

## Archive

- [AI models released in 2026](https://theaiengineering.com/releases/2026): 400 releases
- [AI models released in 2025](https://theaiengineering.com/releases/2025): 422 releases
- [AI models released in 2024](https://theaiengineering.com/releases/2024): 233 releases
- [AI models released in 2023](https://theaiengineering.com/releases/2023): 112 releases

## Cite and reuse this data

These release records are free to reuse under CC BY 4.0, including commercially. Credit The AI Engineering and link to this page's URL. Dates are the labs' official announcement dates, and every record links to its source. List fields in the CSV are separated by " | ".

The AI Engineering Editors. "New AI models: the LLM release tracker." The AI Engineering, updated Nov 16, 2026. https://theaiengineering.com/releases

HTML link:

```html
<a href="https://theaiengineering.com/releases">New AI models: the LLM release tracker</a> (The AI Engineering)
```

- Title: New AI models: the LLM release tracker
- Author: The AI Engineering Editors
- Publisher: The AI Engineering
- Updated: 2026-11-16
- URL: https://theaiengineering.com/releases
- Markdown: https://theaiengineering.com/releases.md
- License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/)
- Full dataset as CSV: https://theaiengineering.com/releases.csv
- Full dataset as JSON: https://theaiengineering.com/releases.json
