AI models released in August 2025

38 releases from 23 labs in August 2025, including LongCat-Flash-Chat, Step-Audio 2 mini and Command A Translate. Every entry is dated to the lab's official announcement and links to it.

Releases
38
Labs
23
Open weights
26

August 2025

38 releases
  1. MeituanLLMOpen weights

    LongCat-Flash-Chat

    LongCat-Flash-Chat is Meituan's first LongCat model, a non-thinking mixture-of-experts foundation model with a dynamic computation mechanism that varies activated parameters per token, aimed at efficient inference and agentic tasks.

    Modalities
    text → text
    Context
    125K context
    Parameters
    MoE: 560B total, 18.6B to 31.3B active
    License
    MIT
    At launch
    API, Consumer app, Open weights
    • 560B total parameters, activating 18.6B to 31.3B (about 27B on average) per token
    • Zero-computation experts and a shortcut-connected MoE architecture
    • 128K context at launch; the API platform opened on September 5, 2025
    • Weights released under MIT
    Source: Change Log | LongCat API Docs
  2. StepFunSpeechOpen weights

    Step-Audio 2 mini

    Step-Audio 2 mini is the open-weight member of StepFun's Step-Audio 2 family, an end-to-end multimodal model for audio understanding and speech conversation. A base version shipped alongside it, and a Think variant followed in September.

    Modalities
    audio, text → text, audio
    License
    Apache-2.0
    At launch
    Open weights
    • Step-Audio 2 technical report was published July 23, 2025
    • Step-Audio 2 mini and Step-Audio 2 mini Base released August 29, 2025
    • Step-Audio 2 mini Think added September 15, 2025
    • Released under Apache-2.0
    Source: GitHub - stepfun-ai/Step-Audio2
  3. CohereLLMOpen weights

    Command A Translate

    Translation-specialized model in Cohere's Command A family covering 23 business languages, offered with Deep Translation, an agentic multi-step approach that refines translations further.

    Modalities
    text → text
    Parameters
    111B
    License
    CC-BY-NC-4.0
    At launch
    API, Open weights
    • Supports 8K input and 8K output tokens
    • Deployable on one H100 or A100 with 4-bit quantization
    • Research weights released as command-a-translate-08-2025
    Source: Command A Translate: Secure Enterprise Translation | Cohere
  4. OpenAISpeechProprietary

    gpt-realtime

    gpt-realtime is a speech-to-speech model released with the general availability of the Realtime API. OpenAI reports gains in following complex instructions, precise tool calling, and more natural, expressive speech.

    Modalities
    text, audio, image → text, audio
    Context
    32K context
    At launch
    API
    • API model ID gpt-realtime (snapshot gpt-realtime-2025-08-28)
    • Realtime API added image input, remote MCP servers, and SIP phone calling
    • Two new voices, Cedar and Marin, exclusive to the Realtime API
    • 32,000-token context window per the API model page
    Source: Introducing gpt-realtime and Realtime API updates for production voice agents | OpenAI
  5. xAICodingProprietary

    Grok Code Fast 1

    Grok Code Fast 1 is a fast, economical reasoning model for agentic coding, built from scratch on a new architecture with a programming-heavy pretraining corpus and post-training on real-world pull requests and coding tasks.

    Modalities
    text → text
    At launch
    API
    • API model ID: grok-code-fast-1 (released earlier in stealth as sonic)
    • Free for a limited time via launch partners including GitHub Copilot and Cursor
    • Retired May 15, 2026; requests route to grok-build-0.1
    Source: Grok Code Fast 1 | SpaceXSI
  6. TencentMultimodalOpen weights

    HunyuanVideo-Foley

    HunyuanVideo-Foley is an open text-video-to-audio model that generates sound effects and audio matched to video content, aimed at short video creation, film production, advertising, and game development.

    Modalities
    text, video → audio
    License
    Tencent Hunyuan Community License Agreement
    At launch
    Open weights
    • Self-developed 48kHz audio VAE for high-fidelity output
    • A smaller XL model with offload inference was added September 29, 2025
    Source: GitHub - Tencent-Hunyuan/HunyuanVideo-Foley: HunyuanVideo-Foley: Multimodal Diffusion with Representation Alignment for High-Fidelity Foley Audio Generation.
  7. KuaishouMultimodalOpen weights

    Keye-VL-1.5-8B

    Keye-VL-1.5 is an upgraded 8B multimodal model from Kuaishou's Keye team that improves short-video comprehension with a Slow-Fast video encoding strategy and new long chain-of-thought and reinforcement learning stages.

    Modalities
    text, image, video → text
    Parameters
    8B
    License
    Apache-2.0
    At launch
    Open weights
    • Slow-Fast video encoding strategy for video understanding
    • Progressive four-stage pre-training
    • Weights released under Apache-2.0
    Source: GitHub - Kwai-Keye/Keye
  8. MicrosoftLLMProprietary

    MAI-1-preview

    MAI-1-preview is Microsoft AI's first in-house foundation language model, a mixture-of-experts model released for public testing on LMArena and rolled out for certain text use cases in Copilot.

    Modalities
    text → text
    At launch
    Consumer app
    • Mixture-of-experts model pre-trained and post-trained on about 15,000 NVIDIA H100 GPUs
    • Public testing on LMArena, with API access for trusted testers
    • Rolled out to certain Copilot text use cases
    Source: Two in-house models in support of our mission
  9. MicrosoftSpeechProprietary

    MAI-Voice-1

    MAI-Voice-1 is Microsoft AI's first in-house speech generation model, producing expressive audio for single and multi-speaker scenarios. It launched in Copilot Daily and Podcasts and later reached Microsoft Foundry.

    Modalities
    text → audio
    At launch
    Consumer app, API
    • Generates a minute of audio in under a second on a single GPU, per Microsoft
    • Launched in Copilot Daily, Copilot Podcasts and Copilot Labs
    • Added to Microsoft Foundry with custom voice creation on April 2, 2026
    Source: Two in-house models in support of our mission
  10. Ai2SpeechOpen weights

    OLMoASR

    Family of fully open English automatic speech recognition models from Ai2, trained from scratch on a curated, large-scale dataset and released in sizes from 244M to 1.5B.

    Modalities
    audio → text
    Parameters
    244M to 1.5B
    License
    Apache-2.0
    At launch
    Open weights
    • Sizes include small.en at 244M, medium.en at 769M and two 1.5B large.en models
    • The large.en-v2 model was trained on 680,000 hours of audio
    • Released under Apache 2.0
    Source: OLMoASR: A series of open speech recognition models | Ai2
  11. GoogleImageProprietary

    Gemini 2.5 Flash Image (Nano Banana)

    Gemini 2.5 Flash Image, nicknamed Nano Banana, is Google's image generation and editing model that can blend multiple images, keep characters consistent and make targeted edits with natural language.

    Modalities
    text, image → text, image
    Context
    64K context
    At launch
    API, Cloud platforms
    • Preview model gemini-2.5-flash-image-preview in the Gemini API, Google AI Studio and Vertex AI
    • Generally available on Oct 2, 2025
    • All generated or edited images carry an invisible SynthID watermark
    Source: Introducing Gemini 2.5 Flash Image, our state-of-the-art image model
  12. Nous ResearchReasoningOpen weights

    Hermes 4

    Hybrid-mode reasoning models from Nous Research released with a technical report, led by a frontier model based on Llama 3.1 405B.

    Modalities
    text → text
    Parameters
    14B, 70B and 405B
    License
    Llama 3 Community License (70B, 405B); Apache-2.0 (14B)
    At launch
    Open weights
    • Released as Hermes-4-Llama-3.1-405B, Hermes-4-Llama-3.1-70B and Hermes-4-14B
    • 14B model is a small dense variant for local inference
    • Hermes 4 Technical Report published the same day
    Source: Releases | Nous Research
  13. Alibaba QwenVideoOpen weights

    Wan2.2-S2V-14B

    Wan2.2-S2V-14B is an audio-driven cinematic video generation model in the Wan2.2 family that generates video from speech and a reference image.

    Modalities
    text, image, audio → video
    Parameters
    14B
    License
    Apache-2.0
    At launch
    Open weights
    • Supports 480P and 720P output
    • Text-to-speech support through CosyVoice added September 5, 2025
    • Released under the Apache 2.0 license
    Source: Wan2.2 (GitHub)
  14. xAILLMOpen weights

    Grok 2 (open weights)

    xAI published the weights of Grok 2, a post-trained model it trained and used in 2024, on Hugging Face, with instructions for serving the roughly 500 GB checkpoint with SGLang across eight GPUs.

    Modalities
    text → text
    License
    Grok 2 Community License Agreement
    At launch
    Open weights
    • Hugging Face repo: xai-org/grok-2 (42 files, about 500 GB)
    • Config lists a mixture of 8 experts with 2 active per token
    • Checkpoint is TP=8 and needs 8 GPUs with more than 40 GB memory each
    • LICENSE file was retitled xAI Community License Agreement on November 4, 2025
    Source: xai-org/grok-2 · Hugging Face
  15. CohereReasoningOpen weights

    Command A Reasoning

    Cohere's first reasoning model, built for enterprise agentic workflows with a user-set token budget that switches between maximum accuracy and higher throughput.

    Modalities
    text → text
    Context
    250K context
    Parameters
    111B
    License
    CC-BY-NC-4.0
    At launch
    API, Open weights
    • Runs on a single H100 or A100 at 128K context
    • Core generative model behind the North agent platform
    • Research weights released as command-a-reasoning-08-2025
    Source: Command A Reasoning: Control for AI Agents | Cohere
  16. DeepSeekLLMOpen weights

    DeepSeek-V3.1

    DeepSeek-V3.1 is a hybrid model that supports thinking and non-thinking modes in one model, with stronger tool use and multi-step agent skills from post-training. DeepSeek called it its first step toward the agent era.

    Modalities
    text → text
    Context
    125K context
    Parameters
    MoE: 671B total, 37B active
    License
    MIT
    At launch
    API, Consumer app, Open weights
    • V3.1 Base adds 840B tokens of continued pre-training for long-context extension on top of V3
    • 128K context for both deepseek-chat and deepseek-reasoner
    • API added Anthropic format support and strict function calling in beta
    Source: DeepSeek-V3.1 Release
  17. ByteDance SeedLLMOpen weights

    Seed-OSS-36B

    Seed-OSS-36B is an open-weights language model series from ByteDance Seed, trained on 12T tokens with native 512K context, a user-controllable thinking budget, and tuning for reasoning and agent tasks.

    Modalities
    text → text
    Context
    500K context
    Parameters
    36B
    License
    Apache-2.0
    At launch
    Open weights
    • Released as Base, Base-woSyn (no synthetic instruction data), and Instruct
    • Native 512K-token context
    • Thinking budget can be adjusted flexibly at inference
    • Weights on Hugging Face under Apache-2.0
    Source: ByteDance-Seed/Seed-OSS-36B-Instruct · Hugging Face
  18. Alibaba QwenImageOpen weights

    Qwen-Image-Edit

    Qwen-Image-Edit is an image editing model built on the 20B Qwen-Image model, extending its text rendering to precise text editing in images. It feeds the input image to Qwen2.5-VL and a VAE encoder for both semantic and appearance editing.

    Modalities
    text, image → image
    Parameters
    20B
    License
    Apache-2.0
    At launch
    Consumer app, Open weights
    • Available through the Image Editing feature on Qwen Chat
    • Weights on Hugging Face and ModelScope
    • Released under the Apache 2.0 license
    Source: Qwen-Image-Edit: Image Editing with Higher Quality and Efficiency
  19. NVIDIAReasoningOpen weights

    NVIDIA Nemotron Nano 2

    9B language model trained from scratch by NVIDIA as a unified model for reasoning and non-reasoning tasks, generating a reasoning trace before its final answer.

    Modalities
    text → text
    Parameters
    9B
    License
    NVIDIA Open Model License
    At launch
    Open weights
    • Supports English, German, Spanish, French, Italian and Japanese
    • A 12B v2 variant was released August 29, 2025
    • Released under the NVIDIA Open Model License
    Source: nvidia/NVIDIA-Nemotron-Nano-9B-v2 · Hugging Face
  20. IBMEmbeddingOpen weights

    Granite Embedding R2

    Second-generation English Granite embedding models based on ModernBERT, with a 149M model replacing the 125M version and a smaller 47M model.

    Modalities
    text → embedding
    Context
    8K context
    Parameters
    149M and 47M
    License
    Apache-2.0
    At launch
    Open weights
    • 8,192-token context length
    • 149M model outputs 768-dimensional vectors; 47M outputs 384
    • Released under Apache 2.0
    Source: ibm-granite/granite-embedding-english-r2 · Hugging Face
  21. GoogleSmall modelOpen weights

    Gemma 3 270M

    Gemma 3 270M is a compact Gemma 3 model designed as a base for task-specific fine-tuning, such as classification and entity extraction. Google calls it the most power-efficient Gemma model to date.

    Modalities
    text → text
    Context
    32K context
    Parameters
    270M
    License
    Gemma Terms of Use
    At launch
    Open weights, Cloud platforms
    • 170M embedding parameters and 100M transformer parameters, with a 256K-token vocabulary
    • Downloads on Hugging Face, Ollama, Kaggle, LM Studio and Docker
    • Pretrained, instruction-tuned and QAT checkpoints released
    Source: Introducing Gemma 3 270M: The compact model for hyper-efficient AI
  22. TencentVideoOpen weights

    Hunyuan-GameCraft-1.0

    Hunyuan-GameCraft is an open framework for high-dynamic interactive game video generation that maps keyboard and mouse inputs into a shared camera representation for fine-grained action control and long video extension.

    Modalities
    text, image → video
    License
    Tencent Hunyuan Community License Agreement
    At launch
    Open weights
    • Generates from a reference image, a text prompt, and keyboard or mouse signals
    • Gradio demo code followed on August 21, 2025
    Source: GitHub - Tencent-Hunyuan/Hunyuan-GameCraft-1.0: Hunyuan-GameCraft: High-dynamic Interactive Game Video Generation with Hybrid History Condition
  23. StepFunImageOpen weights

    NextStep-1

    NextStep-1 is a 14B-parameter autoregressive image generation model that works directly with continuous image tokens rather than discrete ones, released with text-to-image and image editing checkpoints.

    Modalities
    text, image → image
    Parameters
    14B
    License
    Apache-2.0
    At launch
    Open weights
    • Released as NextStep-1-Large-Pretrain, NextStep-1-Large, and NextStep-1-Large-Edit
    • Autoregressive generation over continuous image tokens
    • Weights on Hugging Face under Apache-2.0
    Source: GitHub - stepfun-ai/NextStep-1
  24. NVIDIASpeechOpen weights

    Parakeet TDT 0.6B v3 and Canary-1B v2

    Multilingual speech models covering 25 European languages: Parakeet TDT 0.6B v3 for high-throughput transcription and Canary-1B v2 for transcription and translation.

    Modalities
    audio → text
    Parameters
    600M and 1B
    License
    CC-BY-4.0
    At launch
    Open weights
    • Canary-1B v2 translates between English and 24 languages
    • Parakeet v3 extends the English-only v2
    • Both released under CC BY 4.0
    Source: nvidia/canary-1b-v2 · Hugging Face
  25. Liquid AIMultimodalOpen weights

    LFM2-VL (450M, 1.6B)

    Liquid AI's first vision-language models, extending LFM2 with SigLIP2 NaFlex image encoders to process text and images at native resolution, released as open-weight 450M and 1.6B variants for on-device use.

    Modalities
    text, image → text
    Context
    32K context
    Parameters
    450M; 1.6B
    License
    LFM Open License v1.0
    At launch
    Open weights
    • Native image resolution up to 512x512; larger images split into 512x512 patches
    • Built on LFM2-350M (450M model) and LFM2-1.2B (1.6B model) backbones
    • User-tunable image token count trades speed for quality at inference
    • Weights on Hugging Face; supported by LEAP for edge deployment
    Source: LFM2-VL: Efficient Vision-Language Models
  26. Mistral AIMultimodalProprietary

    Mistral Medium 3.1

    August 2025 update of Mistral Medium 3, Mistral's frontier-class multimodal model, with improvements to tone and performance and the same 128K-token context window.

    Modalities
    text, image → text
    Context
    125K context
    At launch
    API
    • API model ID mistral-medium-2508
    • 128K-token context window
    • Deprecated May 22, 2026 in favor of Mistral Medium 3.5
    Source: Mistral Medium 3.1 | Mistral Docs
  27. Ai2MultimodalOpen weights

    MolmoAct

    Ai2's first Action Reasoning Model, a robotics model that reasons about its environment in 3D before acting, released with an open post-training dataset.

    Modalities
    text, image → action
    Parameters
    7B
    License
    Apache-2.0
    At launch
    Open weights
    • Post-trained on about 10,000 distinct robot episodes
    • Pretrained on a curated subset of Open-X Embodiment data
    • Released under Apache 2.0
    Source: MolmoAct: An Action Reasoning Model that reasons in 3D space | Ai2
  28. Z.aiMultimodalOpen weights

    GLM-4.5V

    GLM-4.5V is Z.ai's open-weights vision-language reasoning model built on the GLM-4.5-Air text model. It covers image, video, and document understanding, GUI agent tasks, and front-end replication from images.

    Modalities
    text, image, video, file → text
    Parameters
    MoE: 106B total, 12B active
    License
    MIT
    At launch
    API, Open weights
    • Adds a thinking mode switch for fast responses or deep reasoning
    • Continues the technical approach of GLM-4.1V-Thinking
    • Weights under the MIT license
    Source: GLM-V (GitHub)
  29. OpenAILLMProprietary

    GPT-5 (with GPT-5 mini and GPT-5 nano)

    GPT-5 is OpenAI's flagship model family, launched in ChatGPT as a unified system that routes between a fast model and GPT-5 thinking, and in the API as gpt-5, gpt-5-mini, and gpt-5-nano with configurable reasoning effort.

    Modalities
    text, image → text
    Context
    400K context
    At launch
    Consumer app, API, Enterprise
    • API models gpt-5, gpt-5-mini, and gpt-5-nano; ChatGPT's non-reasoning model is gpt-5-chat-latest
    • API context totals 400,000 tokens (272,000 input plus 128,000 reasoning and output)
    • GPT-5 pro for ChatGPT Pro users; gpt-5-pro reached the API on October 6, 2025
    • API added a minimal reasoning effort and custom tool calls
    Source: Introducing GPT-5 | OpenAI
  30. MiniMaxSpeechProprietary

    MiniMax Speech 2.5

    MiniMax Speech 2.5 is a text-to-speech model that builds on Speech 02 with stronger multilingual expressiveness, more accurate voice cloning, and language support expanded to more than 40 languages.

    Modalities
    text → audio
    At launch
    API, Consumer app
    • Voice cloning preserves accents and speech habits of the source voice
    • Expanded language support to over 40 languages
    Source: MiniMax Speech 2.5 Launches: Enhanced Multilingual Expressiveness Exceptional Voice Cloning Fidelity
  31. AnthropicLLMProprietary

    Claude Opus 4.1

    Claude Opus 4.1 is an incremental upgrade to Claude Opus 4 with improvements in agentic tasks, real-world coding, and reasoning, including better detail tracking and agentic search for research and data analysis.

    Modalities
    text, image → text
    At launch
    API, Consumer app, Cloud platforms
    • API model ID claude-opus-4-1-20250805
    • Available to paid Claude users, in Claude Code, and on the Claude API, Amazon Bedrock, and Vertex AI
    • Retired on the Claude API on August 5, 2026
    Source: Claude Opus 4.1
  32. ElevenLabsMusicProprietary

    Eleven Music

    ElevenLabs' first music model, generating studio-grade songs from natural language prompts with control over genre, style and structure, with or without vocals, in multiple languages.

    Modalities
    text → audio
    At launch
    Consumer app, API
    • Section-level editing of sound and lyrics
    • Available in the API from August 18, 2025
    • Model ID music_v1, since superseded by Music v2
    Source: Eleven Music is here
  33. OpenAIReasoningOpen weights

    gpt-oss-120b and gpt-oss-20b

    gpt-oss-120b and gpt-oss-20b are OpenAI's open-weight mixture-of-experts reasoning models, released under Apache 2.0 with adjustable reasoning effort, full chain-of-thought, tool use, and Structured Outputs support.

    Modalities
    text → text
    Context
    128K context
    Parameters
    120b: 117B total, 5.1B active; 20b: 21B total, 3.6B active
    License
    Apache-2.0
    At launch
    Open weights, Cloud platforms
    • Natively quantized in MXFP4: 120b runs within 80 GB of memory, 20b within 16 GB
    • Weights on Hugging Face; launch partners included Azure, AWS, vLLM, and Ollama
    • Trained on the harmony response format; o200k_harmony tokenizer open-sourced
    • Native 128k context; API docs list a 131,072-token window
    Source: Introducing gpt-oss | OpenAI
  34. Leonardo.AiImageProprietary

    Lucid Origin

    Leonardo.Ai foundation image model trained for Full HD renders with greater diversity and dynamic range, balancing strong prompt adherence and text rendering with stylistic range.

    Modalities
    text → image
    At launch
    Consumer app
    • Trained for Full HD renders
    • Built by blending training checkpoints in a curated process
    • Offered in Fast and Ultra modes on Leonardo
    Source: Lucid Origin - Leonardo.AI's latest Gen AI model
  35. TencentSmall modelOpen weights

    Hunyuan 0.5B, 1.8B, 4B, and 7B

    Tencent open-sourced four dense Hunyuan language models at 0.5B, 1.8B, 4B, and 7B parameters, each with pretrained and instruct versions, trained with strategies similar to Hunyuan-A13B for deployment from edge devices to servers.

    Modalities
    text → text
    Context
    250K context
    Parameters
    0.5B, 1.8B, 4B, 7B
    License
    Tencent Hunyuan Community License Agreement
    At launch
    Open weights
    • Hybrid fast and slow thinking modes
    • Native 256K context window
    • FP8, GPTQ-Int4, and AWQ-Int4 quantized versions published
    Source: LICENSE · tencent/Hunyuan-7B-Instruct at main
  36. Alibaba QwenImageOpen weights

    Qwen-Image

    Qwen-Image is a 20B MMDiT image foundation model with significant advances in complex text rendering and precise image editing. It renders both alphabetic and logographic text, including English and Chinese.

    Modalities
    text → image
    Parameters
    20B
    License
    Apache-2.0
    At launch
    Consumer app, Open weights
    • Weights on Hugging Face and ModelScope at launch
    • Image Generation available on Qwen Chat
    • Released under the Apache 2.0 license
    Source: Qwen-Image: Crafting with Native Text Rendering
  37. GoogleReasoningProprietary

    Gemini 2.5 Deep Think

    Gemini 2.5 Deep Think is an enhanced reasoning mode for Gemini 2.5 that uses parallel thinking to explore several ideas at once before answering, rolled out in the Gemini app to Google AI Ultra subscribers.

    Modalities
    text → text
    At launch
    Consumer app
    • First previewed at Google I/O on May 20, 2025 as an experimental mode for trusted testers
    • Turned on with a Deep Think toggle in the Gemini app prompt bar
    • A research version was shared with a small group of mathematicians and academics
    Source: Gemini 2.5: Deep Think is now rolling out
  38. IBMLLMOpen weights

    Granite Guardian 3.3

    Granite 3.3 8B-based safety model that judges whether prompts and responses meet specified criteria, with built-in checks for jailbreaks, harm and RAG groundedness.

    Modalities
    text → text
    Parameters
    8B
    License
    Apache-2.0
    At launch
    Open weights
    • Based on Granite 3.3 8B
    • Supports custom criteria beyond pre-baked detectors
    • Released under Apache 2.0
    Source: ibm-granite/granite-guardian-3.3-8b · Hugging Face

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 2025." The AI Engineering, updated Aug 29, 2025. https://theaiengineering.com/releases/2025/08

HTML link

<a href="https://theaiengineering.com/releases/2025/08">AI models released in August 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

Archive