AI models released in May 2025
31 releases from 18 labs in May 2025, including MiMo-VL-7B, FLUX.1 Kontext [pro] and [max] and Codestral Embed. Every entry is dated to the lab's official announcement and links to it.
- Releases
- 31
- Labs
- 18
- Open weights
- 16
May 2025
31 releasesMiMo-VL-7B
MiMo-VL-7B is Xiaomi's open-source vision-language model series, released as SFT and RL checkpoints, with the RL version trained through Mixed On-policy Reinforcement Learning across perception, grounding, reasoning, and preference rewards.
- Modalities
- text, image, video → text
- Parameters
- 7B
- License
- MIT
- At launch
- Open weights
- Four-stage pre-training ending in long-context supervised fine-tuning
- MiMo-VL-7B-2508 update in August 2025 added a switch to turn reasoning off
- Weights under the MIT license
FLUX.1 Kontext [pro] and [max]
In-context image generation and editing models that accept text and reference images, enabling character consistency, local edits and style reference, plus multi-turn iterative editing; [max] adds stronger prompt adherence and typography.
- Modalities
- text, image → image
- At launch
- API, Consumer app, Cloud platforms
- Unifies text-to-image generation and instruction-based editing
- Launched with the BFL Playground
- Available via the BFL API and partners such as fal, Replicate and Together AI
Codestral Embed
Mistral's first embedding model specialized for code, built for high-performance retrieval and semantic understanding over real-world code data in coding assistants and agents.
- Modalities
- text → embedding
- Context
- 8K context
- At launch
- API
- Outputs embeddings at multiple dimensions and precisions
- Dimensions are ordered by relevance, so vectors can be truncated
- API model ID codestral-embed-2505
DeepSeek-R1-0528
DeepSeek-R1-0528 is an update to DeepSeek-R1 with improved reasoning, enhanced front-end generation, and reduced hallucinations. It added JSON output and function calling support to the reasoning model.
- Modalities
- text → text
- License
- MIT
- At launch
- API, Consumer app, Open weights
- Replaced R1 behind the deepseek-reasoner API model with no change to API usage
- Released with DeepSeek-R1-0528-Qwen3-8B, a distilled 8B model
- Weights on Hugging Face under the MIT License
HunyuanVideo-Avatar
HunyuanVideo-Avatar is an open audio-driven human animation model based on a multimodal diffusion transformer that turns avatar images into dynamic, emotion-controllable videos, including multi-character dialogue.
- Modalities
- image, audio → video
- License
- Tencent Hunyuan Community License Agreement
- At launch
- Open weights
- Supports photorealistic, cartoon, 3D-rendered, and anthropomorphic characters
- Portrait, upper-body, and full-body generation
Kling 2.1
Kling 2.1 is a model series that adds Standard and High Quality video generation modes alongside a premium Kling 2.1 Master edition with stronger motion and prompt responsiveness.
- Modalities
- text, image → video
- At launch
- Consumer app
- Standard (720p) and High Quality (1080p) modes
- Kling 2.1 Master offers improved motion and semantic responsiveness
- Start and end frame control for Kling 2.1 arrived on August 21, 2025
Claude Opus 4 and Claude Sonnet 4
Claude Opus 4 and Claude Sonnet 4 are Anthropic's fourth-generation hybrid models, offering near-instant responses or extended thinking, with Opus 4 aimed at long-running coding and agent workflows and Sonnet 4 succeeding Claude 3.7 Sonnet.
- Modalities
- text, image → text
- At launch
- API, Consumer app, Cloud platforms, Enterprise
- API model IDs claude-opus-4-20250514 and claude-sonnet-4-20250514
- Extended thinking with tool use in beta; both models can use tools in parallel
- Opus 4 deployed under the ASL-3 Standard and Sonnet 4 under ASL-2
- Sonnet 4 gained a 1M-token context window in beta on August 12, 2025
Mistral OCR 2
Second-generation Mistral OCR service released together with document annotations, powering Mistral's Document AI stack so users can extract interleaved text and images from documents.
- Modalities
- text, image → text
- At launch
- API
- API model ID mistral-ocr-2505
- Released with the annotations feature
- Deprecated February 27, 2026
Devstral
Agentic coding model built with All Hands AI and trained to resolve real GitHub issues inside agent scaffolds such as OpenHands and SWE-Agent.
- Modalities
- text → text
- Context
- 125K context
- Parameters
- 24B
- License
- Apache-2.0
- At launch
- Open weights, API
- Light enough for a single RTX 4090 or a Mac with 32GB of RAM
- Fine-tuned from Mistral Small 3.1
- API model ID devstral-small-2505
BAGEL
BAGEL is an open-source unified multimodal model from ByteDance Seed that handles understanding, image generation, and image editing in one Mixture-of-Transformer-Experts architecture pretrained on interleaved text, image, video, and web data.
- Modalities
- text, image → text, image
- Parameters
- MoT: 14B total, 7B active
- License
- Apache-2.0
- At launch
- Open weights
- 7B active and 14B total parameters
- Fine-tuned from Qwen2.5-7B-Instruct and SigLIP
- Website, demo, weights, and report released on May 20, 2025
Gemini 2.5 Flash Native Audio
Gemini 2.5 Flash native audio models brought native audio output to the Live API, enabling real-time spoken dialogue with natural voices, proactive audio that ignores background speech, and support for 24+ languages.
- Modalities
- audio, video, text → audio, text
- At launch
- API
- Preview models gemini-2.5-flash-preview-native-audio-dialog and gemini-2.5-flash-exp-native-audio-thinking-dialog
- Announced at Google I/O 2025 with the Gemini 2.5 updates
- Updated native audio previews followed in September and December 2025
Gemini 2.5 Flash TTS and Gemini 2.5 Pro TTS
Gemini 2.5 Flash TTS and Gemini 2.5 Pro TTS are controllable text-to-speech preview models that generate speech with one or two speakers, letting developers direct style, tone and pacing with natural language.
- Modalities
- text → audio
- Context
- 8K context
- At launch
- API
- Pro TTS targets quality on complex prompts; Flash TTS targets cost-efficient everyday applications
- Input token limit of 8,192 per the Gemini API model pages
- Enhanced in December 2025 with improved expressivity, pacing and dialogue
Gemma 3n
Gemma 3n is a mobile-first open model that accepts text, image, audio and video input on devices with limited memory. It was previewed in May 2025 and fully released on June 26, 2025.
- Modalities
- text, image, audio, video → text
- Context
- 32K context
- Parameters
- E2B, E4B (5B and 8B raw)
- License
- Gemma Terms of Use
- At launch
- Open weights, API, Cloud platforms
- Per-Layer Embeddings let E2B and E4B run in about 2GB and 3GB of accelerator memory
- MatFormer architecture nests an E2B sub-model inside E4B
- Text support for 140 languages and multimodal understanding in 35
- Architecture shared with the next generation of Gemini Nano
Imagen 4
Imagen 4 is Google DeepMind's text-to-image model announced at I/O 2025, with output up to 2K resolution and much better spelling and typography. It later shipped in Standard, Ultra and Fast variants.
- Modalities
- text → image
- At launch
- Consumer app, API, Cloud platforms
- Launched in the Gemini app, Whisk, Vertex AI and Workspace apps such as Slides, Vids and Docs
- Imagen 4 and Imagen 4 Ultra entered preview in the Gemini API on June 24, 2025
- Imagen 4 Standard, Ultra and Fast generally available in the Gemini API and Vertex AI from August 14, 2025
MedGemma
MedGemma is a collection of Gemma 3-based open models for medical text and image comprehension, released through Google's Health AI Developer Foundations. The initial release included a 4B multimodal model and a 27B text-only model.
- Modalities
- text, image → text
- Context
- 128K context
- Parameters
- 4B (multimodal), 27B (text)
- License
- Health AI Developer Foundations Terms of Use
- At launch
- Open weights, Cloud platforms
- 4B model ships in pre-trained and instruction-tuned versions
- Supports at least 128K tokens of input and up to 8,192 output tokens
- Gated on Hugging Face under the Health AI Developer Foundations terms
- Intended as a starting point that requires validation before clinical use
Veo 3
Veo 3 is Google DeepMind's video generation model announced at I/O 2025 and the first Veo model to generate audio with video, including dialogue and background sound. It launched in the Gemini app, Flow and Vertex AI.
- Modalities
- text, image → video, audio
- At launch
- Consumer app, API, Cloud platforms
- Initially for U.S. Google AI Ultra subscribers in the Gemini app and Flow, and enterprise users on Vertex AI
- Veo 3 Fast entered preview in the Gemini API on July 31, 2025
- Veo 3 and Veo 3 Fast generally available on Vertex AI from July 29, 2025 and in the Gemini API from September 9, 2025
codex-1 and codex-mini-latest
codex-1 is a version of OpenAI o3 optimized for software engineering that powers the Codex cloud agent research preview. OpenAI also released codex-mini-latest, a version of o4-mini built for the Codex CLI.
- Modalities
- text, image → text
- At launch
- Consumer app, Enterprise, API
- Codex rolled out first to ChatGPT Pro, Enterprise, and Business users; Plus followed June 3, 2025
- codex-mini-latest became the default Codex CLI model and an API model
- codex-1 was trained with reinforcement learning on real-world coding tasks
- codex-mini-latest has a 200,000-token context window per the API model page
Stable Audio Open Small
341M-parameter text-to-audio model developed with Arm and optimized to run entirely on Arm CPUs, generating up to 11 seconds of audio on a smartphone in under 8 seconds.
- Modalities
- text → audio
- Parameters
- 341M
- License
- Stability AI Community License
- At launch
- Open weights
- 341M parameters versus 1.1B for Stable Audio Open
- Optimized with Arm KleidiAI
- Free for commercial use under the Stability AI Community License
Wan2.1-VACE
Wan2.1-VACE is an all-in-one model for video creation and editing in the Wan2.1 family, released with inference code, weights, and a technical report.
- Modalities
- text, image, video → video
- Parameters
- 1.3B, 14B
- License
- Apache-2.0
- At launch
- Open weights
- Released in 1.3B and 14B sizes
- Handles multiple video creation and editing tasks in one model
- Released under the Apache 2.0 license
Seed1.5-VL
Seed1.5-VL is ByteDance Seed's vision-language model, pairing a 532M-parameter SeedViT encoder with a mixture-of-experts language model of 20B active parameters for image, video, GUI, and game understanding and reasoning.
- Modalities
- text, image, video → text
- Parameters
- MoE LLM with 20B active, plus 532M vision encoder
- At launch
- API
- Pretrained on more than 3T tokens of multimodal data
- API on Volcano Engine as Doubao-1.5-thinking-vision-pro
- Technical report and sample code published
Step1X-3D
Step1X-3D is an open framework for generating textured 3D assets, pairing a geometry generation model with a texture synthesis model, released with training code and a curated 3D dataset.
- Modalities
- image → 3d
- License
- Apache-2.0
- At launch
- Open weights
- 1.3B-parameter geometry model and 3.5B-parameter texture model
- Training code and 800K curated 3D asset IDs released
- Weights on Hugging Face under Apache-2.0
Seed1.5-Embedding
Seed1.5-Embedding is a text embedding model further trained from Seed1.5 (Doubao-1.5-pro), optimized for general retrieval and reasoning-intensive search, using a mixture-of-experts backbone with few active parameters and multiple vector dimensions.
- Modalities
- text → text
- At launch
- API
- Two-stage contrastive training on curated data
- Supports multiple embedding dimensions via MRL training
- API available through Volcano Ark as doubao-embedding-large-text
HunyuanCustom
HunyuanCustom is an open multimodal-driven architecture for customized video generation built on HunyuanVideo, starting with single-subject customization and later adding audio-driven and video-driven customization.
- Modalities
- text, image → video
- License
- Tencent Hunyuan Community License Agreement
- At launch
- Open weights
- Audio-driven and video-driven weights released June 6, 2025
- Runs on a single GPU with 8GB VRAM through community tooling
Seed-Coder
Seed-Coder is an open-weights family of 8B code models with base, instruct, and reasoning variants, pretrained on 6T tokens of code data curated largely by an LLM-driven filtering pipeline.
- Modalities
- text → text
- Context
- 32K context
- Parameters
- 8B
- License
- MIT
- At launch
- Open weights
- Base and Instruct support 32K context; Reasoning supports 64K
- Released under the MIT License
- A BF16 version of the reasoning model was added later
Mistral Medium 3
Enterprise-focused multimodal model that Mistral positioned as balancing strong performance with lower cost and simpler deployment, with particular strength in coding and multimodal understanding.
- Modalities
- text, image → text
- Context
- 125K context
- At launch
- API, Enterprise
- 128K-token context window
- Deployable on any cloud or self-hosted on four or more GPUs
- API model ID mistral-medium-2505
Gemini 2.5 Pro Preview (I/O edition)
Gemini 2.5 Pro Preview (I/O edition) was an updated 2.5 Pro released ahead of Google I/O 2025 with significantly improved coding capabilities, especially for building interactive web apps.
- Modalities
- text, image, audio, video → text
- At launch
- API, Cloud platforms, Consumer app
- Model ID gemini-2.5-pro-preview-05-06; the 03-25 preview automatically pointed to the new version
- Available in the Gemini API via Google AI Studio and Vertex AI and in the Gemini app
- Further updated as gemini-2.5-pro-preview-06-05 before general availability
LTX-Video 13B 0.9.7
13B-parameter LTX-Video model with improved prompt adherence and physical understanding and a multi-scale rendering pipeline, released with FP8 quantized weights and spatial and temporal upscalers.
- Modalities
- text, image → video
- Parameters
- 13B
- License
- LTXV Open Weights License 0.X
- At launch
- Open weights
- 13B parameters
- Distilled 13B model followed on May 14, 2025
- Licensed under the LTXV Open Weights License
Granite 4.0 Tiny Preview
Early preview of the Granite 4.0 architecture, a 7B fine-grained hybrid mixture-of-experts instruct model released ahead of the full Granite 4.0 family.
- Modalities
- text → text
- Parameters
- MoE: 7B total
- License
- Apache-2.0
- At launch
- Open weights
- Fine-grained hybrid MoE architecture
- Fine-tuned from Granite-4.0-Tiny-Base-Preview
- Released under Apache 2.0
xGen-small
Compact enterprise-focused language models trained from scratch in 4B and 9B sizes, combining domain-focused data curation, length extension, instruction tuning and reinforcement learning for long-context performance at low cost.
- Modalities
- text → text
- Context
- 125K context
- Parameters
- 4B and 9B
- License
- CC-BY-NC-4.0
- At launch
- Open weights
- Trained on a budget of eight trillion tokens
- Base and instruct variants supported 128k-token context
- Released for research use under CC BY-NC 4.0
Parakeet TDT 0.6B v2
600-million-parameter English automatic speech recognition model from NVIDIA with punctuation, capitalization and accurate timestamp prediction, released openly under CC BY 4.0.
- Modalities
- audio → text
- Parameters
- 600M
- License
- CC-BY-4.0
- At launch
- Open weights
- Supports English transcription
- Predicts word-level timestamps accurately
- Succeeded by the multilingual v3 in August 2025
Suno v4.5
More expressive Suno model with more dynamic music, greater genre variety and accuracy, and richer vocals, released with tools to combine Personas, Covers and Extend.
- Modalities
- text, audio → audio
- At launch
- Consumer app
- Greater variety and accuracy in genres
- Richer vocals
- Longer Covers and combined features
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 2025." The AI Engineering, updated May 30, 2025. https://theaiengineering.com/releases/2025/05
HTML link
<a href="https://theaiengineering.com/releases/2025/05">AI models released in May 2025</a> (The AI Engineering)- Full dataset
- 1,167 releases from 55 labs
- Coverage
- to
- Updated
- License
- CC BY 4.0
- Publisher
- The AI Engineering