AI models released in May 2026
29 releases from 21 labs in May 2026, including Cosmos 3, Grok Build 0.1 and Step 3.7 Flash. Every entry is dated to the lab's official announcement and links to it.
- Releases
- 29
- Labs
- 21
- Open weights
- 12
May 2026
29 releasesCosmos 3
Omnimodal world models that generate video, images, audio and robot action commands from text, image, video and action trajectory inputs for physical AI.
- Modalities
- text, image, video → video, image, audio
- Parameters
- Nano 16B; Super 64B
- License
- OpenMDW-1.1
- At launch
- Open weights
- Released on Hugging Face and GitHub
- Cosmos3-Edge, a 4B model, followed on July 20, 2026
- Released under the OpenMDW 1.1 license
Grok Build 0.1
Grok Build 0.1 is SpaceXSI's coding model trained for agentic coding tasks such as web development, debugging, and MCP support. It is the same model that powers the Grok Build coding agent, released on the API in public beta.
- Modalities
- text, image → text
- Context
- 250K context
- At launch
- API
- API model ID: grok-build-0.1
- 256k-token context window per the docs model page
- Served at 100+ tokens per second
- Docs list grok-code-fast-1 as an alias of grok-build-0.1
Step 3.7 Flash
Step 3.7 Flash is a sparse mixture-of-experts vision-language model that builds on Step 3.5 Flash, adding native visual input for agent workflows that combine perception, search, coding, and reasoning at high throughput.
- Modalities
- text, image, video → text
- Context
- 250K context
- Parameters
- MoE: 198B total, about 11B active
- License
- Apache-2.0
- At launch
- API, Open weights
- 196B-parameter language backbone plus a 1.8B-parameter vision encoder
- Activates about 11B parameters per token with a 256K context window
- Three selectable reasoning levels: low, medium, and high
- Weights released under Apache-2.0
Claude Opus 4.8
Claude Opus 4.8 is an upgrade to Claude Opus 4.7 with improvements in coding, agentic tasks, reasoning, and practical knowledge work. It launched alongside effort controls in claude.ai and a dynamic workflows research preview in Claude Code.
- Modalities
- text, image → text
- Context
- 1M context
- At launch
- API, Consumer app, Cloud platforms
- API model ID claude-opus-4-8
- 1M-token context window by default on the Claude API, Amazon Bedrock, Google Cloud, and Microsoft Foundry
- 128K max output tokens
- Supports mid-conversation system messages and high-resolution image input
LFM2.5-8B-A1B
An updated on-device mixture-of-experts model built for tool calling, expanding LFM2-8B-A1B with a 128K context window, pretraining scaled from 12T to 38T tokens, a doubled vocabulary, and reasoning-only output.
- Modalities
- text → text
- Context
- 125K context
- Parameters
- MoE: 8.3B total, 1.5B active
- License
- LFM Open License v1.0
- At launch
- Consumer app, Open weights
- Context window expanded from 32,768 to 128,000 tokens
- Vocabulary grows from 65,536 to 128,000 tokens for non-Latin scripts
- Produces an explicit chain of thought before every final answer
- Base and post-trained models on Hugging Face; testable on Liquid Playground
Music v2
Second-generation ElevenLabs music model with better vocals, instrumentation and arrangement across genres and improved multilingual support, powering ElevenMusic, ElevenAPI and ElevenCreative.
- Modalities
- text, audio → audio
- At launch
- Consumer app, API
- Model ID music_v2
- Generates from prompts, composition plans and earlier songs
- Reference-based sound control added later
Keye-VL-2.0-30B-A3B
Keye-VL-2.0-30B-A3B is the 30B-class flagship of Kuaishou's Keye multimodal series, using a DSA sparse-attention architecture to support long-video understanding over a 256K context.
- Modalities
- text, image, video → text
- Context
- 250K context
- Parameters
- MoE: 30B total, 3B active
- License
- Apache-2.0
- At launch
- Open weights
- Sparse attention architecture for hour-long video understanding
- Nearly lossless reasoning over 256K context, per the model card
- Weights released under Apache-2.0
Aleph 2.0
Upgrade to Runway's flagship video editing model, launched with Edit Studio, that edits clips up to 30 seconds at 1080p with localized changes and image-level control from an edited frame.
- Modalities
- text, image, video → video
- At launch
- Consumer app
- Edits clips up to 30 seconds at 1080p
- Applies a user-edited frame across the whole video
- Added to Runway MCP in July 2026
Fara1.5
Fara1.5 is Microsoft's second generation of open computer use agent models, released in 4B, 9B and 27B sizes and fine-tuned from Qwen3.5 on synthetic web-task trajectories generated by the FaraGen1.5 pipeline.
- Modalities
- text, image → text
- Context
- 256K context
- Parameters
- 4B, 9B, 27B
- License
- MIT
- At launch
- Open weights, API
- Released in 4B, 9B and 27B sizes with a 262,144-token context length
- Supervised fine-tuned from Qwen3.5 base models
- 27B version also offered on Microsoft Foundry
Hy-MT2
Hy-MT2 is a family of open fast-thinking multilingual translation models in 1.8B, 7B, and 30B-A3B mixture-of-experts sizes, supporting translation among 33 languages and following translation instructions given in multiple languages.
- Modalities
- text → text
- Parameters
- 1.8B, 7B, and MoE: 30B total, 3B active
- License
- Apache-2.0
- At launch
- Open weights
- Three sizes: 1.8B, 7B, and 30B-A3B (MoE)
- Released with IFMTBench, a benchmark for translation instruction following
- FP8 and GGUF versions published on Hugging Face
LongCat-Video-Avatar 1.5
LongCat-Video-Avatar 1.5 is an upgraded open-source framework for audio-driven human video generation from Meituan's LongCat team, building on the LongCat-Video foundation model.
- Modalities
- text, image, audio → video
- License
- MIT
- At launch
- Open weights
- Generates talking and expressive character video from audio
- Supports single- and multi-speaker audio-to-video generation
- Weights released under MIT
Command A+
Final model in the Command A family and Cohere's first mixture-of-experts model, combining vision input, reasoning, translation and agentic tasks in one model.
- Modalities
- text, image → text
- Context
- 125K context
- Parameters
- MoE: 218B total, 25B active
- License
- Apache-2.0
- At launch
- API, Open weights, Enterprise
- Supports 48 languages, including all official EU languages
- Deployable on one B200 or two H100 GPUs
- Supports up to 64K output tokens
Qwen3.7-Max
Qwen3.7-Max is Qwen's proprietary model designed for the agent era, targeting coding, office workflow automation, and long autonomous runs. Qwen demonstrated a 35-hour autonomous kernel optimization run with more than 1,000 tool calls.
- Modalities
- text → text
- At launch
- API
- Built to work across scaffolds such as Claude Code, OpenClaw, and Qwen Code
- Supports MCP integrations and multi-agent orchestration
- Available through Alibaba Cloud Model Studio
Stable Audio 3.0 Large
Largest Stable Audio 3.0 model with the family's most advanced musicality, built for music platforms and creative applications that need low-latency, high-volume generation.
- Modalities
- text, audio → audio
- At launch
- API, Enterprise
- Generates tracks of more than six minutes
- Supports single and multi-segment audio inpainting and continuation
- Available via the Stability AI API and enterprise self-hosting
Stable Audio 3.0 Small, Small SFX and Medium
Open-weight members of the Stable Audio 3.0 family, trained on fully licensed data with a semantic-acoustic autoencoder; Small runs full music composition on-device and Medium generates tracks over six minutes.
- Modalities
- text, audio → audio
- License
- Stability AI Community License
- At launch
- Open weights
- 3.0 Small generates up to two minutes; Medium more than six minutes
- Small SFX targets on-device sound effects
- LoRA training documentation published with the weights
Gemini 3.5 Flash
Gemini 3.5 Flash was the first model in the Gemini 3.5 family, which Google built to combine frontier intelligence with action, targeting agents, coding and complex long-horizon tasks at Flash speed.
- Modalities
- text, image, audio, video → text
- Context
- 1M context
- At launch
- API, Cloud platforms, Consumer app
- Launched at Google I/O 2026 in the Gemini app, AI Mode in Search, Google Antigravity and the Gemini API
- Released directly as the generally available gemini-3.5-flash
- Computer use support added in public preview on June 24, 2026
- Replaced by Gemini 3.6 Flash, with gemini-3.5-flash requests routed to 3.6 Flash from Oct 8, 2026
Gemini Omni Flash
Gemini Omni Flash is the first model in Google's Omni family, which combines Gemini's reasoning with generation, starting with video it can create and edit conversationally from text, image, video and audio inputs.
- Modalities
- text, image, video, audio → video, audio, text
- At launch
- API, Cloud platforms, Consumer app
- Rolled out first to the Gemini app, Google Flow and YouTube Shorts at Google I/O 2026
- Public preview in the Gemini API and Google AI Studio (gemini-omni-flash-preview) began June 30, 2026
- Only voice references were supported for audio input at launch
Recraft V4.1
Update to Recraft V4 with more natural photorealism, new illustration styles and better aesthetic results from short prompts, plus a Utility variant for simpler outputs.
- Modalities
- text → image
- At launch
- Consumer app, API
- New Utility variant
- V4.1 Vector for illustrations and text
- Pro variant also offered
ERNIE 5.1
ERNIE 5.1 is Baidu's successor to ERNIE 5.0, derived through an elastic training framework that compresses total parameters to about one-third and active parameters to about one-half of ERNIE 5.0.
- Modalities
- text → text
- At launch
- Consumer app
- Four-stage post-training that includes multi-teacher on-policy distillation and online reinforcement learning
- Available on the ERNIE website and in an ERNIE 5.1 playground on Baidu AI Studio
EMO
Research mixture-of-experts model pretrained so that modular structure emerges from data, letting a task use about 12.5% of its experts with near full performance.
- Modalities
- text → text
- Parameters
- MoE: 14B total, 1B active
- License
- Apache-2.0
- At launch
- Open weights
- 128 experts in total with 8 active
- Trained on 1 trillion tokens
- Released under Apache 2.0
StepAudio 2.5 Realtime
StepAudio 2.5 Realtime is StepFun's real-time voice conversation model, focused on understanding and expressing paralinguistic cues, user-defined personas set through the API, and stable role-play behavior.
- Modalities
- text, audio → text, audio
- At launch
- API
- Interprets tone, pace, pauses, sighs, and laughter in user speech
- Personality, background, preferences, and conversational boundaries can be set through the API
- Trained on more than 10,000 native personas expanded into a million-scale persona feature set
GPT-Realtime-2
GPT-Realtime-2 is a speech-to-speech model for the Realtime API with GPT-5-class reasoning and adjustable reasoning effort, aimed at voice agents that call tools, handle interruptions, and sustain longer sessions.
- Modalities
- text, audio, image → text, audio
- Context
- 125K context
- At launch
- API
- Context window raised from 32K to 128K tokens compared with GPT-Realtime-1.5
- Reasoning effort levels minimal, low, medium, high, and xhigh, with low as the default
- Supports preambles and parallel tool calls with audible tool transparency
- API model ID gpt-realtime-2; testable in the Playground
GPT-Realtime-Translate
GPT-Realtime-Translate is a streaming speech-to-speech translation model that translates live speech from more than 70 input languages into 13 output languages, returning translated audio and transcripts while the speaker is still talking.
- Modalities
- audio → audio, text
- Context
- 16K context
- At launch
- API
- API model ID gpt-realtime-translate on a dedicated realtime translation endpoint (v1/realtime/translations)
- 16,000-token context window per the API model page
- Launched alongside GPT-Realtime-2 and GPT-Realtime-Whisper
GPT-Realtime-Whisper
GPT-Realtime-Whisper is a streaming speech-to-text model for the Realtime API that transcribes audio as people speak, built for live captions, meeting notes, and voice agents that need low-latency transcripts.
- Modalities
- audio, text → text
- Context
- 16K context
- At launch
- API
- API model ID gpt-realtime-whisper on the Realtime transcription sessions endpoint
- Lets developers tune the tradeoff between latency and accuracy
- 16,000-token context window per the API model page
Grok Imagine Quality Mode
Grok Imagine Quality Mode is an image generation and editing model on the Grok Imagine API, served as grok-imagine-image-quality, with higher realism, cleaner multilingual text rendering, and tighter prompt following.
- Modalities
- text, image → image
- At launch
- API
- Retired
- API model ID: grok-imagine-image-quality
- Launched for enterprise developers and teams via the Grok Imagine API
- grok-imagine-image-pro requests redirected to it from May 15, 2026
- Retired November 2, 2026; requests are now served by grok-imagine-image-2.0
xAI retired the grok-imagine-image-quality model on November 2, 2026, after a 60-day notice that began on September 2, and began serving its requests with grok-imagine-image-2.0 at low quality.
Source: Grok Imagine Quality Mode API | SpaceXSIRetirement: grok-imagine-image-quality Retirement on November 2, 2026
GPT-5.5 Instant
GPT-5.5 Instant replaced GPT-5.3 Instant as ChatGPT's default model for all users, with improved factuality, clearer and more concise answers, better analysis of image uploads, and more personalized responses.
- Modalities
- text, image → text
- At launch
- Consumer app, API
- Available in the API as chat-latest
- GPT-5.3 Instant stayed available to paid users for three months before retirement
- Updated on May 28, 2026 to improve response style and quality
Granite Switch 4.1
Modular models that compose a Granite 4.1 base model with adapters activated by control tokens, so one endpoint can switch between chat, decisions and guardrail checks.
- Modalities
- text → text
- Parameters
- 3B, 8B and 30B
- License
- Apache-2.0
- At launch
- Open weights
- Released first as 3B, 8B and 30B previews
- Non-preview 8B and 30B versions released October 14, 2026
- Revision 4.1.1 on October 27, 2026 added embedded speech-to-text
MolmoAct 2
Upgrade to Ai2's Action Reasoning Model that pairs the Molmo 2-ER embodied reasoning model with a flow-matching action expert for real-world robots.
- Modalities
- text, image → action
- At launch
- Open weights
- Molmo 2-ER released as the reasoning backbone
- Includes an open action tokenizer
- Full code and training data released later
Uni-1.1
Luma's Unified Intelligence image model, which runs reasoning and image generation in the same model, opened to developers as the Uni-1.1 REST API for generation and natural-language editing.
- Modalities
- text, image → image
- At launch
- API, Consumer app
- Up to nine reference images per request
- Generate Image and Modify Image endpoints
- Builds on Uni-1, which Luma launched the previous month
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 May 2026." The AI Engineering, updated May 31, 2026. https://theaiengineering.com/releases/2026/05
HTML link
<a href="https://theaiengineering.com/releases/2026/05">AI models released in May 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