AI models released in April 2026
45 releases from 21 labs in April 2026, including Granite 4.1, Granite Embedding Multilingual R2 and Granite Speech 4.1. Every entry is dated to the lab's official announcement and links to it.
- Releases
- 45
- Labs
- 21
- Open weights
- 22
April 2026
45 releasesGranite 4.1
Dense Granite 4.1 language models with an improved post-training pipeline, released as base and instruct checkpoints in 3B, 8B and 30B sizes.
- Modalities
- text → text
- Context
- 128K context
- Parameters
- 3B, 8B and 30B
- License
- Apache-2.0
- At launch
- Open weights
- Supports 12 languages including Arabic, Japanese and Chinese
- Granite Guardian 4.1 8B, fine-tuned from Granite 4.1 8B, was also released in April 2026
- Released under Apache 2.0
Granite Embedding Multilingual R2
Multilingual Granite embedding models rebuilt on ModernBERT, supporting 200+ languages and code retrieval with a context window extended from 512 to 32,768 tokens.
- Modalities
- text → embedding
- Context
- 32K context
- Parameters
- 311M and 97M
- License
- Apache-2.0
- At launch
- Open weights
- 311M model outputs 768-dimensional vectors
- Matryoshka truncation down to 128 dimensions
- Released under Apache 2.0
Granite Speech 4.1
Speech-language model for multilingual speech recognition and translation in English, French, German, Spanish, Portuguese and Japanese, with punctuation and truecasing.
- Modalities
- audio → text
- Parameters
- 2B
- License
- Apache-2.0
- At launch
- Open weights
- Plus variant adds speaker-attributed ASR and word timestamps
- NAR variant uses a non-autoregressive architecture
- Released under Apache 2.0
Granite Vision 4.1
Vision-language model built on Granite 4.1 3B for structured document extraction, covering chart conversion, complex table extraction and semantic key-value pair extraction.
- Modalities
- text, image → text
- Parameters
- 4B: 3.4B LLM + 0.6B vision
- License
- Apache-2.0
- At launch
- Open weights
- Chart extraction to CSV, summaries and code
- Trained with the ChartNet dataset
- Released under Apache 2.0
Step Image Edit 2
Step Image Edit 2 is a lightweight 3.5B-parameter StepFun model for image generation and editing, supporting Chinese and English text rendering, local edits, subject consistency, and style transfer with fast response times.
- Modalities
- text, image → image
- Parameters
- 3.5B
- At launch
- API
- Generates or edits an image in about 0.5 to 2 seconds
- Launched on the StepFun open platform and Step Plan
- StepFun retired the model and its image APIs on October 10, 2026
Mistral Medium 3.5
Dense 128B flagship that merges instruction following, reasoning and coding in one set of weights, with per-request reasoning effort and a vision encoder trained from scratch.
- Modalities
- text, image → text
- Context
- 250K context
- Parameters
- 128B
- License
- Modified MIT License
- At launch
- Open weights, API, Consumer app
- Showcased May 22, 2026 as the default model for remote agents in Vibe and Le Chat
- Self-hostable on as few as four GPUs
- API model ID mistral-medium-3-5
Nemotron 3 Nano Omni
Omnimodal reasoning model built on the Nemotron 3 Nano 30B hybrid Mamba-Transformer backbone that understands video, audio, images and text for Q&A, transcription and document intelligence.
- Modalities
- text, image, video, audio → text
- Context
- 250K context
- Parameters
- MoE: 31B total, about 3B active
- License
- NVIDIA Open Model Agreement
- At launch
- Open weights
- Maximum context of 256K tokens
- Activates about 3B parameters per token
- Released under the NVIDIA Open Model Agreement
DeepSeek-V4 (Preview)
DeepSeek-V4 is a preview release of two open-weights mixture-of-experts models, V4-Pro and V4-Flash. It introduced a hybrid attention design combining token-wise compression with DeepSeek Sparse Attention and made 1M-token context the default across DeepSeek services.
- Modalities
- text → text
- Context
- 1M context
- Parameters
- MoE: 1.6T total, 49B active (V4-Pro); 284B total, 13B active (V4-Flash)
- License
- MIT
- At launch
- API, Consumer app, Open weights
- Both models pre-trained on more than 32T tokens and support thinking and non-thinking modes
- API supports OpenAI ChatCompletions and Anthropic formats via deepseek-v4-pro and deepseek-v4-flash
- Legacy deepseek-chat and deepseek-reasoner names were scheduled for retirement on July 24, 2026
StepAudio 2.5 ASR
StepAudio 2.5 ASR is StepFun's speech recognition model that brings Step 3.5 Flash's multi-token prediction to transcription and reuses a 32K language model context to transcribe up to 30 minutes of audio in one pass.
- Modalities
- audio → text
- Context
- 32K context
- At launch
- API
- ASR plus MTP-5 architecture with peak inference of 500 tokens per second
- Transcribes up to 30 minutes of audio in a single request without slicing
- Launched on the StepFun open platform and Step Plan
GPT-5.5 (with GPT-5.5 Pro)
GPT-5.5 is a frontier model for complex professional work such as coding, research, data analysis, and operating software, which OpenAI says matches GPT-5.4 per-token latency while using fewer tokens on Codex tasks.
- Modalities
- text, image → text
- Context
- 1.1M context
- At launch
- Consumer app, API, Enterprise
- GPT-5.5 Thinking for Plus, Pro, Business, and Enterprise; GPT-5.5 Pro for Pro, Business, and Enterprise
- API models gpt-5.5 and gpt-5.5-pro became available on April 24, 2026
- 1,050,000-token context window per the API model page; 400K context in Codex
- Reasoning effort defaults to medium in the API
Grok Voice 1.0
Grok Voice 1.0 is SpaceXSI's flagship voice model for the Speech to Speech API, built for complex, multi-step voice agent workflows in customer support and sales that need precise data entry and heavy tool calling.
- Modalities
- audio → audio
- At launch
- API
- Available via the Speech to Speech (voice agent) API
- Natively supports 25+ languages
- Built to reason through edge cases before responding
- Developed with partners including Starlink for noisy telephony conditions
Hy3 preview
Hy3 preview is the first model trained on Tencent's rebuilt Hy infrastructure, an open mixture-of-experts model with three reasoning modes that improves complex reasoning, instruction following, context learning, coding, and agent tasks.
- Modalities
- text → text
- Context
- 250K context
- Parameters
- MoE: 295B total, 21B active
- License
- Tencent HY Community License Agreement
- At launch
- Open weights, API
- 295B total parameters with 21B active and a 256K context window
- Weights on GitHub, Hugging Face, ModelScope, and GitCode with vLLM and SGLang support
- API on Tencent Cloud, plus Token Plans for individuals
MiMo-V2.5-ASR
MiMo-V2.5-ASR is Xiaomi's open-source speech recognition model for bilingual Chinese and English, Chinese dialects, code-switching, lyrics transcription, knowledge-heavy content, and noisy or multi-speaker audio.
- Modalities
- audio → text
- License
- MIT
- At launch
- API, Open weights
- Open-sourced with the MiMo-V2.5-TTS series launch on April 23, 2026
- Supports Wu, Cantonese, Hokkien, Sichuanese, and other dialects
- Added to the MiMo API Platform on June 2, 2026
- Weights under the MIT license
MiMo-V2.5-Pro
MiMo-V2.5-Pro is Xiaomi's flagship agentic language model, improving on MiMo-V2-Pro in general agent tasks, complex software engineering, and long-horizon work spanning more than a thousand tool calls.
- Modalities
- text → text
- Context
- 1M context
- Parameters
- MoE: 1.02T total, 42B active
- License
- MIT
- At launch
- API, Open weights
- Public beta on the MiMo API Platform began April 23, 2026
- Weights open-sourced under the MIT license on April 27, 2026
- Hybrid attention with a 6:1 sliding-window to global ratio and three MTP modules
MiMo-V2.5-TTS Series
MiMo-V2.5-TTS is a family of three Xiaomi speech synthesis models covering premium stock voices, voice design from a text description, and voice cloning, all steerable through style instructions and inline audio tags.
- Modalities
- text → audio
- At launch
- API, Consumer app
- Models: mimo-v2.5-tts, mimo-v2.5-tts-voicedesign, and mimo-v2.5-tts-voiceclone
- Released on the MiMo API Platform on April 23, 2026, per its model release log
- Accepts director-script style input describing characters, scenes, and per-line instructions
Seed3D 2.0
Seed3D 2.0 is ByteDance Seed's second-generation image-to-3D model, with an upgraded architecture for higher geometric precision and material quality so generated assets come closer to production use in embodied AI and manufacturing.
- Modalities
- image → 3d
- At launch
- API
- Uses a larger DiT for coarse geometry generation
- Adopts a mixture-of-experts design for high-resolution material generation
- API available on Volcano Engine Ark as Doubao-Seed3D-2.0
- Technical report published alongside the release
MiMo-V2.5
MiMo-V2.5 is a natively omnimodal Xiaomi model that adds in-house visual and audio encoders to the MiMo-V2-Flash backbone, built for agentic tasks with up to 1M tokens of context.
- Modalities
- text, image, video, audio → text
- Context
- 1M context
- Parameters
- MoE: 310B total, 15B active
- License
- MIT
- At launch
- API, Open weights
- Trained on about 48T tokens
- Xiaomi reports agentic performance above MiMo-V2-Pro
- Weights, including the Base model, released under MIT on April 27, 2026
- Public beta on the MiMo API Platform from April 23, 2026
OpenAI Privacy Filter
OpenAI Privacy Filter is an open-weight model for detecting and redacting personally identifiable information in text, built as a bidirectional token classifier that labels input in a single pass and can run locally.
- Modalities
- text → text
- Context
- 125K context
- Parameters
- 1.5B total, 50M active
- License
- Apache-2.0
- At launch
- Open weights
- 1.5B total parameters with 50M active parameters
- Supports up to 128,000 tokens of context
- Released under Apache 2.0 on Hugging Face (openai/privacy-filter) and GitHub
- Predicts PII spans across eight privacy categories
Qwen3.6-27B
Qwen3.6-27B is a dense 27B open-weights multimodal model focused on agentic coding. Qwen says it surpasses the previous open flagship, Qwen3.5-397B-A17B, on major coding benchmarks while being simpler to deploy.
- Modalities
- text, image → text
- Context
- 256K context
- Parameters
- 27B
- License
- Apache-2.0
- At launch
- Consumer app, Open weights
- Supports multimodal thinking and non-thinking modes
- Native 262,144-token context, extensible to 1,010,000 tokens
- Available on Qwen Studio; weights under the Apache 2.0 license
GPT Image 2 (ChatGPT Images 2.0)
ChatGPT Images 2.0 introduced GPT Image 2, OpenAI's new image generation and editing model for ChatGPT and the API, adding flexible image sizes, high-fidelity image inputs, and images with thinking for paid ChatGPT plans.
- Modalities
- text, image → image
- At launch
- Consumer app, API
- Available on all ChatGPT plans; images with thinking on paid plans with Thinking and Pro models
- API model ID gpt-image-2 (snapshot gpt-image-2-2026-04-21) with Batch API support
- Transparent backgrounds added in preview for gpt-image-2 on August 20, 2026
Kimi K2.6
Kimi K2.6 is Moonshot AI's open-weights model focused on long-horizon coding, long-running execution, and agent swarm capabilities. It improves on Kimi K2.5 across programming languages and tasks such as front-end, DevOps, and performance optimization.
- Modalities
- text, image, video → text
- Context
- 250K context
- Parameters
- MoE: 1T total, 32B active
- License
- Modified MIT License
- At launch
- API, Consumer app, Open weights
- Available on Kimi.com, the Kimi App, the API, and Kimi Code
- Native multimodal model with the 400M-parameter MoonViT vision encoder
- Weights and code under the Modified MIT License
Qwen3.6-Max-Preview
Qwen3.6-Max-Preview is an early preview of Qwen's next proprietary flagship, released after Qwen3.6-Plus. It brings stronger world knowledge, instruction following, and agentic coding than Qwen3.6-Plus.
- Modalities
- text → text
- At launch
- API, Consumer app
- Available on Qwen Studio
- API model name qwen3.6-max-preview on Alibaba Cloud Model Studio
- Qwen describes it as still under active development
Grok Speech to Text
Grok Speech to Text is a standalone transcription API built on the stack behind Grok Voice, offering batch REST and real-time WebSocket modes with word-level timestamps, speaker diarization, and multichannel support.
- Modalities
- audio → text
- At launch
- API
- Supports 25+ languages
- Inverse Text Normalization formats numbers, dates, and currencies
- Docs release notes list general availability on April 15, 2026
- Docs later refer to it as grok-voice-transcribe-1.0
Grok Text to Speech
Grok Text to Speech is a standalone speech synthesis API built on the stack behind Grok Voice, with REST and real-time WebSocket modes and inline speech tags for laughter, whispers, emphasis, and pauses.
- Modalities
- text → audio
- At launch
- API
- Docs release notes list general availability on March 16, 2026
- Announced alongside Grok Speech to Text
- All voices can speak every supported language; Eve is the default voice
- Output formats range from MP3 to telephony-optimized mu-law
Claude Opus 4.7
Claude Opus 4.7 is an upgrade to Claude Opus 4.6 focused on advanced software engineering, higher-resolution vision, and instruction following. It is the first model Anthropic shipped with new safeguards that automatically detect and block prohibited or high-risk cybersecurity uses.
- Modalities
- text, image → text
- Context
- 1M context
- At launch
- API, Consumer app, Cloud platforms
- API model ID claude-opus-4-7; 1M-token context window and 128K max output tokens
- Updated tokenizer can map the same input to roughly 1.0 to 1.35 times as many tokens
- Maximum image resolution raised to 2,576 pixels on the long edge
- Adds an xhigh effort level and task budgets in beta
GPT-Rosalind
GPT-Rosalind is OpenAI's frontier reasoning model for biology, drug discovery, and translational medicine, and the first model in its GPT-Rosalind life sciences series. It launched through a trusted access program for qualified Enterprise customers in the U.S.
- At launch
- API, Enterprise
- Available in ChatGPT, Codex, and the API for qualified customers through trusted access
- Launched alongside a free Life Sciences research plugin for Codex that connects to more than 50 scientific tools and data sources
- Launch customers included Amgen, Moderna, the Allen Institute, and Thermo Fisher Scientific
- Left research preview on September 11, 2026
HY-World 2.0
HY-World 2.0 is an open multi-modal world model framework that generates and reconstructs navigable 3D worlds from text, single images, multi-view images, or videos, producing meshes or Gaussian splatting representations.
- Modalities
- text, image, video → 3d
- License
- Tencent HY-World 2.0 Community License Agreement
- At launch
- Open weights
- WorldMirror 2.0 weights and the technical report released April 16, 2026
- HY-Pano 2.0 and WorldStereo 2.0 weights followed in May 2026
- Follows HunyuanWorld 1.0, 1.1 (WorldMirror), and 1.5 (WorldPlay)
StepAudio 2.5 TTS
StepAudio 2.5 TTS is StepFun's speech generation model with passage-level and in-text control over emotion, tone, rhythm, and pauses, plus zero-shot voice cloning with adjustable emotion and speaking style.
- Modalities
- text, audio → audio
- At launch
- API
- Global context control sets the emotional tone, character state, and scene for a whole passage
- Controls can be given as natural-language instructions
- Launched on the StepFun open platform and Step Plan
ERNIE-Image
ERNIE-Image is Baidu's open text-to-image model, built on a single-stream Diffusion Transformer with a lightweight Prompt Enhancer, and released alongside a distilled ERNIE-Image-Turbo variant for faster generation.
- Modalities
- text → image
- Parameters
- 8B
- License
- Apache-2.0
- At launch
- Open weights
- 8B DiT parameters; runs on consumer GPUs with 24GB of VRAM
- ERNIE-Image uses about 50 inference steps; ERNIE-Image-Turbo uses 8 steps after DMD and RL optimization
- Targets dense text rendering, posters, comics, and multi-panel layouts
Gemini 3.1 Flash TTS
Gemini 3.1 Flash TTS is a cost-efficient, expressive text-to-speech model that adds granular audio tags for directing vocal style, pace and delivery with natural language across more than 70 languages.
- Modalities
- text → audio
- Context
- 8K context
- At launch
- API, Cloud platforms
- Preview in the Gemini API and Google AI Studio, and for enterprises on Vertex AI
- All generated audio is watermarked with SynthID
- Streaming output support added on June 17, 2026
Qwen3.6-35B-A3B
Qwen3.6-35B-A3B is a sparse open-weights mixture-of-experts model focused on agentic coding, surpassing Qwen3.5-35B-A3B by a wide margin. It supports multimodal thinking and non-thinking modes.
- Modalities
- text, image → text
- Context
- 256K context
- Parameters
- MoE: 35B total, 3B active
- License
- Apache-2.0
- At launch
- API, Consumer app, Open weights
- Offered on the Model Studio API as Qwen3.6-Flash
- Native 262,144-token context, extensible to 1,010,000 tokens
- Weights on Hugging Face and ModelScope under the Apache 2.0 license
Gemini Robotics-ER 1.6
Gemini Robotics-ER 1.6 is an updated embodied reasoning model for robotics with new capabilities such as instrument reading and improved spatial and physical reasoning, available to developers in preview.
- Modalities
- text, image, audio, video → text
- Context
- 128K context
- At launch
- API
- Preview model gemini-robotics-er-1.6-preview in the Gemini API and Google AI Studio
- Replaced gemini-robotics-er-1.5-preview, which shut down on Apr 30, 2026
- Input token limit of 131,072 per the Gemini API model page
MAI-Image-2-Efficient
MAI-Image-2-Efficient is a faster, lower-cost version of Microsoft AI's MAI-Image-2 text-to-image model, aimed at high-volume production work such as product shots, marketing creatives and batch pipelines.
- Modalities
- text → image
- At launch
- API, Consumer app
- Available in Microsoft Foundry and MAI Playground at launch
- Rolling out across Copilot and Bing
- Handles short-form in-image text such as headlines and labels
Midjourney V8.1
Second V8 release with a familiar V7-like aesthetic, a faster HD mode made the default, the return of image prompts and weights, and an updated Describe tool.
- Modalities
- text, image → image
- At launch
- Consumer app
- Became the default model on June 10, 2026
- HD images render at 2K without upscaling
- Standard jobs render about 4 to 5 times faster than earlier versions
MiniMax Music 2.6
MiniMax Music 2.6 updates MiniMax's music generation model with a new Cover mode for reworking existing songs, a reworked low end for stronger bass and drums, and faster first output.
- Modalities
- text, audio → audio
- At launch
- API, Consumer app
- Cover mode is new in Music 2.6
- First-packet latency reduced to under 20 seconds
- Supports original, instrumental, and Cover generation modes through the API
HY-Embodied-0.5
HY-Embodied-0.5 is a suite of embodied foundation models for real-world agents, using a Mixture-of-Transformers architecture for spatial-temporal perception and embodied reasoning. Tencent released the edge-oriented MoT-2B variant as open weights.
- Modalities
- text, image → text
- Parameters
- MoT: 4B total, 2.2B active
- License
- Tencent HY Community License Agreement
- At launch
- Open weights
- MoT-2B has 4B total parameters with 2.2B activated during inference
- A 32B variant is described for complex reasoning; reasoning is distilled into the 2B model
- Released with official inference code
Seeduplex
Seeduplex is a native full-duplex speech model that listens while speaking, replacing the earlier half-duplex Doubao end-to-end voice model to improve conversational pacing, interruption handling, and suppression of background interference.
- Modalities
- audio → audio
- At launch
- Consumer app
- Decides on its own whether to reply, keep listening, or yield to an interruption
- Can draw on background environmental audio as conversational context
- Fully rolled out in the Doubao app's voice call mode
LFM2.5-VL-450M
An updated 450M vision-language model that adds bounding-box object detection, function calling support, and better instruction following over LFM2-VL-450M, with pretraining scaled from 10T to 28T tokens.
- Modalities
- text, image → text
- Context
- 32K context
- Parameters
- 450M
- License
- LFM Open License v1.0
- At launch
- Consumer app, Open weights
- Improved multilingual image understanding across eight languages
- Built on the LFM2.5-350M backbone per model card
- Available on Hugging Face, LEAP, and Liquid Playground
Muse Spark
Muse Spark is the first model from Meta Superintelligence Labs, a natively multimodal reasoning model with tool use, visual chain of thought and multi-agent orchestration that powers Meta AI.
- Modalities
- text, image → text
- At launch
- Consumer app
- Available at launch on meta.ai and in the Meta AI app
- Contemplating mode runs multiple agents that reason in parallel
- API access opened only as a private preview for select users
- Weights were not released; Meta said it hopes to open-source future versions
GLM-5.1
GLM-5.1 is Z.ai's flagship model for agentic engineering, designed for long-horizon tasks. Z.ai says it can work independently for up to 8 hours in a single run, from planning and execution to delivery.
- Modalities
- text → text
- Context
- 200K context
- Parameters
- MoE: 744B total, 40B active
- License
- MIT
- At launch
- API, Open weights
- Trained with multi-turn SFT, RL, and a process-quality evaluation framework
- Up to 128K output tokens
- Weights under the MIT license
HY-OmniWeaving
HY-OmniWeaving is an open unified video generation model from the HunyuanVideo team, built on HunyuanVideo 1.5, that supports free-form multimodal composition and reasoning-augmented video generation in a single model.
- Modalities
- text, image, video → video
- License
- Tencent HY Community License Agreement
- At launch
- Open weights
- Built on the HunyuanVideo 1.5 backbone
- Released with the IntelligentVBench benchmark
- Code and weights released April 3, 2026
Gemma 4
Gemma 4 is Google DeepMind's fourth generation of open models, released in E2B, E4B, 26B mixture-of-experts and 31B dense sizes under Apache 2.0. All sizes accept images and video, and the edge models add audio input.
- Modalities
- text, image, video, audio → text
- Context
- 256K context
- Parameters
- E2B (2.3B effective), E4B (4.5B effective), 26B MoE (3.8B active), 31B
- License
- Apache-2.0
- At launch
- Open weights, API, Cloud platforms
- Context window of 128K tokens for E2B and E4B and 256K for 26B and 31B
- 26B MoE has 25.2B total and 3.8B active parameters
- Native function calling, structured JSON output and system instructions; trained on 140+ languages
- Unquantized 31B fits on a single 80GB NVIDIA H100
GLM-5V-Turbo
GLM-5V-Turbo is Z.ai's multimodal coding foundation model, combining visual understanding with coding ability. It targets GUI agents and coding agents that need to see an environment, plan actions, and execute tasks.
- Modalities
- text, image, video, file → text
- Context
- 200K context
- At launch
- API
- Strengthened fine-grained, geometric, and spatial understanding
- Adds multimodal tools such as box drawing, screenshots, and web page reading
- Up to 128K output tokens
MAI-Transcribe-1
MAI-Transcribe-1 is Microsoft AI's first speech-to-text model, a multilingual transcription model supporting 25 languages and built for noisy, real-world audio such as meetings, call centers and voice agents.
- Modalities
- audio → text
- At launch
- API, Consumer app
- Supports 25 languages
- Available in Microsoft Foundry and MAI Playground
- Launched alongside Foundry availability for MAI-Voice-1 and MAI-Image-2
Qwen3.6-Plus
Qwen3.6-Plus is Qwen's hosted model that succeeded Qwen3.5-Plus, with major upgrades to agentic coding from front-end work to repository-level tasks. It also improves multimodal perception and reasoning.
- Modalities
- text, image → text
- Context
- 1M context
- At launch
- API
- 1M-token context window by default
- Addresses community feedback from the Qwen3.5-Plus deployment
- Available through Alibaba Cloud Model Studio
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 April 2026." The AI Engineering, updated Apr 29, 2026. https://theaiengineering.com/releases/2026/04
HTML link
<a href="https://theaiengineering.com/releases/2026/04">AI models released in April 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