AI models released in August 2026
33 releases from 18 labs in August 2026, including Hy4 preview, Cohere Parse and Gemini Omni 1.1 Flash. Every entry is dated to the lab's official announcement and links to it.
- Releases
- 33
- Labs
- 18
- Open weights
- 17
August 2026
33 releasesHy4 preview
Hy4 preview is an early open-weights version of Tencent's Hy4 flagship, a mixture-of-experts model aimed at long-horizon software engineering, document-heavy office work, game development, and scientific research.
- Modalities
- text → text
- Context
- 1M context
- Parameters
- MoE: 770B total, 49B active
- License
- Apache-2.0
- At launch
- Open weights, API, Consumer app
- 770B total parameters with 49B active and a 1M-token context window
- Weights on Hugging Face, ModelScope, GitHub, and AtomGit
- Available on Tencent Cloud and OpenRouter and in CodeBuddy, WorkBuddy, Yuanbao, and ima
Cohere Parse
Document parsing model that turns complex documents into structured Markdown, extracting text in reading order along with tables, forms, images, captions and layout locations.
- Modalities
- text, image → text
- Context
- 8K context
- Parameters
- 2.3B
- At launch
- API, Cloud platforms, Enterprise
- Model ID parse-v5.0
- Outputs HTML tables, bounding boxes and image descriptions
- Available on Microsoft Foundry, AWS SageMaker and Model Vault
Gemini Omni 1.1 Flash
Gemini Omni 1.1 Flash is the generally available, production-ready update to Gemini Omni Flash, adding video extension, first and last frame interpolation, and output resolution control up to 4K.
- Modalities
- text, image, video → video, audio, text
- Context
- 128K context
- At launch
- API, Cloud platforms
- Model ID gemini-omni-1.1-flash in the Gemini API and Google AI Studio
- Extends scenes using up to 10 seconds of prior context
- Adds 360p previews alongside 720p, 1080p and 4K output, with 1080p and 4K produced by upscaling
Midjourney V8.2 Edit Model
First V8.2 image edit model, supporting instruction-based editing, generation from up to four reference images, inpainting and outpainting, with personalization, moodboards and style references.
- Modalities
- text, image → image
- At launch
- Consumer app
- Replaces Omni Reference, Character Reference and Retexture
- Available on midjourney.com, alpha.midjourney.com and Discord via --edit
- An image quality update to the edit model followed days later
Gemini 3.5 Transcribe
Gemini 3.5 Transcribe is a speech-to-text model built on Gemini's audio understanding, released with a streaming Live API variant for voice agents, live captioning and post-call analytics.
- Modalities
- audio → text
- At launch
- API
- Two models: gemini-3.5-transcribe (non-streaming) and gemini-3.5-transcribe-live (bidirectional streaming)
- Speaker diarization, word-level timestamps and custom vocabulary biasing of up to 1,000 terms
- Google describes it as a major advancement over its previous transcription model, Chirp 3
GLM-5.3-Flash
GLM-5.3-Flash is the first natively multimodal model in Z.ai's GLM-5 series, trained from a new base model. It is the first GLM model to combine sparse and linear attention in a hybrid architecture that cuts long-context serving costs.
- Modalities
- text, image, video, file → text
- Context
- 1M context
- Parameters
- MoE: 320B total, 18B active
- License
- MIT
- At launch
- API, Consumer app, Open weights
- Native visual capabilities let it observe interfaces, rendered output, and interaction feedback
- Up to 128K output tokens
- Extends beyond coding to office document and financial research workflows
- Weights on Hugging Face and ModelScope under the MIT license
Qwen3.8-Flash-Next
Qwen3.8-Flash-Next is an open-weights multimodal mixture-of-experts model that previews the architecture planned for Qwen4. It combines Gated DeltaNet with Qwen Sparse Attention and adds Gated Residual and N-gram Embedding.
- Modalities
- text, image → text
- Context
- 256K context
- Parameters
- MoE: 125B total, 6B active, plus 51B n-gram embedding
- License
- Qwen Community License 1.0
- At launch
- Open weights
- Native 262,144-token context, extensible to 1,000,000 tokens with YaRN
- Qwen says training cost about 1/9 of Qwen3.7-Plus
- Weights on Hugging Face and ModelScope under the Qwen Community License 1.0
Granite 4.2
Granite generation that introduces native chain-of-thought reasoning, with full thinking, non-thinking and low-effort modes in 3B, 8B and 30B models.
- Modalities
- text → text
- Context
- 500K context
- Parameters
- 3B, 8B and 30B
- License
- Apache-2.0
- At launch
- Open weights
- Supports 128K tokens natively, extendable to 512K
- Targets math, coding and agentic tool calling
- Released under Apache 2.0
DeepSeek-V4-Flash-Vision-Exp
DeepSeek-V4-Flash-Vision-Exp is DeepSeek's first experimental multimodal model in the V4 family, adding visual modules to the V4-Flash architecture. DeepSeek says it matches V4-Flash on text tasks while improving multimodal agent performance.
- Modalities
- text, image → text
- License
- MIT
- At launch
- API, Open weights
- API model name deepseek-v4-flash-vision-exp, supporting mixed text and image input
- Launched alongside a new Files API for reusing uploaded images
- Retired on the API on September 10, 2026, when V4.1-Flash replaced it
- Weights published on Hugging Face under the MIT License
Ruby
Native SDR-to-HDR color grading model that converts SDR video from any source into true HDR without re-rendering pixels, delivering HDR10, HLG, ProRes or OpenEXR masters.
- Modalities
- video → video
- At launch
- Consumer app, API
- Accepts SDR clips up to 30 seconds and up to 4K
- Keeps source pixels and audio intact
- Available in Tool Mode, Workflows and Runway Dev for Max and higher plans
GLM-5.3
GLM-5.3 is Z.ai's flagship coding model, which uses the same base model as GLM-5.2 with all gains coming from post-training. Z.ai reports emergent cybersecurity capability in vulnerability discovery and exploitation.
- Modalities
- text → text
- Context
- 1M context
- Parameters
- MoE: 744B total, 40B active
- License
- GLM-5.3 License
- At launch
- API, Consumer app, Open weights
- Text-only input with up to 128K output tokens
- Reasoning is always enabled, with low, high, and max effort levels
- Listed in Z.ai API release notes on August 18, 2026; weights released after safety evaluation
- Weights under the GLM-5.3 License, which adds a security review for very large model-as-a-service providers
Pika Music
Music generation model that turns text, lyrics, a reference voice, an existing track or combinations of these into finished songs through a single workflow.
- Modalities
- text, audio → audio
- At launch
- API
- Announced August 14, 2026 with the Pika Audio model family
- Supports voice conditioning and music references
- Available through the Pika API Club
Pika SFX
Text-to-sound-effects model that generates focused, ready-to-use effects in real time, from single events to timed sequences, while avoiding unrequested speech, music and noise.
- Modalities
- text → audio
- At launch
- API
- Announced August 14, 2026 with the Pika Audio model family
- Handles material, space and perspective cues
- Available through the Pika API Club
Pika Soundtrack
Video-to-audio model in Pika's first audio family that turns video into a synchronized full-scene soundscape of motion-aware sound effects, voice, music and ambience.
- Modalities
- video, text → audio
- At launch
- API
- Announced August 14, 2026 with the Pika Audio model family
- Optional instructions steer what to emphasize or leave out
- Available through the Pika API Club
Pika Speech
Text-to-speech model built as a 3B flow-matching transformer that generates 48 kHz speech in a few denoising steps and clones voices from a few seconds of reference audio.
- Modalities
- text → audio
- Parameters
- 3B
- At launch
- API
- 3B-parameter flow-matching transformer
- Up to five minutes of speech per request
- Controllable style, pace and duration
Qwen3.8-27B
Qwen3.8-27B is a dense 27B open-weights model in the Qwen3.8 series, released on Hugging Face and ModelScope after Qwen3.8-Max. It is a causal language model with a vision encoder.
- Modalities
- text, image → text
- Context
- 256K context
- Parameters
- 27B
- License
- Apache-2.0
- At launch
- Open weights
- Native 262,144-token context, extensible to 1,000,000 tokens
- Released under the Apache 2.0 license
- Follows the Qwen3.8-2.4T-A95B open-weights release on August 12, 2026
DeepSeek-V4-Pro (GA, 0813)
DeepSeek-V4-Pro-0813 is the general availability release of DeepSeek-V4-Pro, superseding the April preview with greatly enhanced agentic capabilities. It keeps the preview's model structure and adds an attached DSpark speculative decoding module.
- Modalities
- text → text
- Context
- 1M context
- Parameters
- MoE: 1.6T total, 49B active
- License
- MIT
- At launch
- API, Consumer app, Open weights
- Rolled out on the DeepSeek app, web, and API under the unchanged deepseek-v4-pro model name
- Added low, high, and max reasoning effort levels for V4-Pro and V4-Flash
- DeepSeek API gained native OpenAI Responses API support with this release
Gemini 3.7 Flash
Gemini 3.7 Flash is a Gemini workhorse model for coding and agents, released three weeks after 3.6 Flash with substantial improvements across software engineering, knowledge work and web development.
- Modalities
- text, image, audio, video → text
- Context
- 1M context
- At launch
- API, Consumer app
- Generally available in the Gemini API via Google AI Studio and Android Studio, and in Google Antigravity
- Powers Gemini Spark, the personal agent in the Gemini app for Google AI Pro and Ultra subscribers
- Replaced by Gemini 3.8 Flash, with requests routed to 3.8 Flash from Oct 8, 2026
Grok 4.6
Grok 4.6 builds on Grok 4.5 with a focus on long-running agents and more ambitious interactive and visual work, such as turning a broad product idea into a working first version of an application.
- Modalities
- text, image → text
- Context
- 500K context
- At launch
- API
- API model ID: grok-4.6, with a 500k-token context window
- Available at launch in Cursor, Grok Build, the API, OpenRouter, Vercel, and Cloudflare
- Reasoning effort settings: low, medium, high (default), and xhigh
- A fast variant is also offered
LFM2.5-VL-3B
An updated 3B vision-language model building on LFM2-VL-3B with improved screen and UI understanding, grounding, function calling for text and image inputs, and multi-image input, answering directly without a reasoning phase.
- Modalities
- text, image → text
- Context
- 32K context
- Parameters
- 3.1B
- License
- LFM Open License v1.0
- At launch
- Consumer app, Open weights
- Built on the LFM2.5-2.6B-Base backbone with a 128,000-token vocabulary
- Vision pretraining scaled by 4x tokens
- 32,768-token context length per model card
- Available on Hugging Face and Liquid Playground
North Micro Vision Instruct
Compact open-weight vision-language model from Cohere with native-resolution image support, intended as a foundation for prototyping, task-specific fine-tuning and specialized multimodal applications.
- Modalities
- text, image → text
- Context
- 125K context
- Parameters
- 2.4B
- License
- Apache-2.0
- At launch
- Open weights
- About 2B language model plus a 400M vision encoder
- Validated for multimodal prompts up to 8K tokens
- Fine-tuning recipes available from NVIDIA and Axolotl
LTX-2.5
LTX model superseding LTX-Video and LTX-2, adding native multi-shot generation, a diffusion video decoder, a custom Gemma 4 12B text encoder and a raw pretrained checkpoint for fine-tuning.
- Modalities
- text, image, video, audio → video, audio
- License
- LTX-2.x Community License
- At launch
- Open weights, API, Consumer app
- Native multi-shot scenes with consistent characters and voices
- Native EXR workflow in ACES and DaVinci Wide Gamut
- Released as open weights and in the LTX API
MAI-Code-1.1-Flash
MAI-Code-1.1-Flash is Microsoft AI's updated small coding model for GitHub Copilot, focused on CLI and .NET tasks and using fewer tokens per task than MAI-Code-1-Flash. A quantized on-device version followed in October 2026.
- Modalities
- text → text
- Context
- 250K context
- At launch
- Consumer app
- In production in GitHub Copilot at launch
- October 7, 2026 update added a 3-bit quantized version that runs locally with a full 256K context window
- Local use recommended on devices with more than 120GB of RAM
Nemotron 3.5 Lightning
Nemotron 3.5 language model built on a hybrid Mamba-2 and mixture-of-experts architecture, released alongside several speculative decoding methods to speed up text generation.
- Modalities
- text → text
- Context
- 1M context
- Parameters
- MoE: 30B total, 3B active
- License
- OpenMDW-1.1
- At launch
- Open weights
- Up to 1M tokens of context, with 256K used for single H100 deployment
- Trained between December 2025 and May 2026
- Released under the OpenMDW 1.1 license
MAI-Image-2.6
MAI-Image-2.6 is Microsoft AI's text-to-image and image editing model succeeding MAI-Image-2.5, with stronger text rendering, portraits, 3D imagery and commercial photorealistic output. Developer access in Microsoft Foundry followed on September 4, 2026.
- Modalities
- text, image → image
- At launch
- API
- Brought to developers in Microsoft Foundry on September 4, 2026
- Generally available in Foundry per an October 27, 2026 update
- Supports multiple image references, web grounding and control over format and resolution
Muse Glimmer
Muse Glimmer is a 30B open-weights model from Meta Superintelligence Labs, distilled from Muse Spark and built for always-on local agent workflows on consumer hardware.
- Modalities
- text, image → text
- Context
- 128K context
- Parameters
- 30B (dense)
- License
- Apache-2.0
- At launch
- Open weights
- Released under Apache 2.0, the most permissive license Meta has used for an open model
- Accepts interleaved text and images through a dedicated perception encoder
- Trained on Muse Spark outputs using logit distillation
- Weights on Hugging Face, including GGUF and ExecuTorch builds
Grok Imagine Image 2.0
Imagine Image 2.0 is SpaceXSI's image generation and editing model, built for close instruction following, designer-style typography and layout, and preserving inputs across generations and edits. It became the Quality Mode on grok.com/imagine.
- Modalities
- text, image → image
- At launch
- Consumer app, API
- API model ID: grok-imagine-image-2.0
- Generally available as Quality Mode on grok.com/imagine and the iOS and Android apps
- Multi-reference editing accepts up to 5 input images per generation
- Adds region editing, segmentation, background removal, and smart resize
Wan-Animate-2
Wan-Animate-2 is an end-to-end character animation model that directly consumes driving videos in a redesigned Diffusion Transformer. It targets high-fidelity motion and strong identity preservation.
- Modalities
- text, image, video → video
- Parameters
- 14B
- License
- Apache-2.0
- At launch
- Open weights
- Released as Base and Distillation weights on August 7, 2026
- Published on Hugging Face as Wan2.2-Animate-2-14B
- Released under the Apache 2.0 license
Muse Spark 1.2
Muse Spark 1.2 is a coding-focused update to Muse Spark 1.1, launched alongside Muse Code, Meta's terminal coding agent, with gains in code generation, debugging and codebase understanding.
- Modalities
- text, image, video, audio → text
- Context
- 1M context
- At launch
- API
- Launched with Muse Code, a beta terminal coding agent
- Available in Muse Code and the Meta Model API with expanded global access
- Meta scaled up training compute on coding tasks
- API docs list a 1,048,576-token context window
SeedRealtime
SeedRealtime is a native audio-visual full-duplex model that jointly understands audio, video, and timing in continuous streams, letting the Doubao app hold real-time video calls in which it watches, listens, and speaks, sometimes proactively.
- Modalities
- audio, video → audio
- At launch
- Consumer app
- Unified architecture natively fuses audio, video, and text
- Decides when to speak without relying on an external voice activity detector
- Can weave tool calls into its spoken responses
- Fully rolled out in the Doubao app's video call mode
LFM2.5-2.6B
An agentic 2.6B model that runs on-device, pre-trained on about 34T tokens with a 128K context extension, a vocabulary doubled to 128K, and a four-stage post-training pipeline ending in agentic reinforcement learning.
- Modalities
- text → text
- Context
- 128K context
- Parameters
- 2.6B
- License
- LFM Open License v1.0
- At launch
- Consumer app, Open weights
- 131,072-token context length per model card
- Post-training: SFT, teacher specialization, multi-domain on-policy distillation, agentic RL
- Base (LFM2.5-2.6B-Base) and post-trained checkpoints on Hugging Face
Shieldstral 1.0
Compact multimodal safety classifier that takes a plain-language policy question at inference time and returns a calibrated safety score for text and images without retraining.
- Modalities
- text, image → text
- Context
- 32K context
- Parameters
- 3.8B
- License
- Apache-2.0
- At launch
- Open weights, API
- Runs on a single 16GB NVIDIA GPU
- Released as an inaugural member of the Open Secure AI Alliance with NVIDIA
- In public preview on Mistral's platform
Qwen3.8-Max
Qwen3.8-Max is the most capable model in Alibaba's Qwen family to date, built on the Qwen3.5 architecture for coding, work, research, and long-horizon tasks. It is the first Qwen-Max-class model with open weights.
- Modalities
- text → text
- Context
- 256K context
- Parameters
- MoE: 2.4T total, 95B active
- License
- Qwen3.8-Max License
- At launch
- API, Open weights
- Weights published as Qwen3.8-2.4T-A95B on Hugging Face and ModelScope on August 12, 2026
- Native 262,144-token context, extensible to about 1M tokens
- API available on QwenCloud at launch
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 " | ".
Citation
The AI Engineering Editors. "AI models released in August 2026." The AI Engineering, updated Aug 28, 2026. https://theaiengineering.com/releases/2026/08
HTML link
<a href="https://theaiengineering.com/releases/2026/08">AI models released in August 2026</a> (The AI Engineering)- Full dataset
- 1,167 releases from 55 labs
- Coverage
- to
- Updated
- License
- CC BY 4.0
- Publisher
- The AI Engineering