AI models released in September 2025

43 releases from 21 labs in September 2025, including GLM-4.6, Sora 2 and Claude Sonnet 4.5. Every entry is dated to the lab's official announcement and links to it.

Releases
43
Labs
21
Open weights
27

September 2025

43 releases
  1. Z.aiLLMOpen weights

    GLM-4.6

    GLM-4.6 is Z.ai's flagship coding model and the successor to GLM-4.5. It expands the context window from 128K to 200K tokens and supports tool use during reasoning.

    Modalities
    text → text
    Context
    200K context
    Parameters
    MoE: 355B total, 32B active
    License
    MIT
    At launch
    API, Open weights
    • Better real-world performance in Claude Code, Cline, Roo Code, and Kilo Code
    • Stronger tool-using and search-based agents
    • Up to 128K output tokens; weights under the MIT license
    Source: GLM-4.6: Advanced Agentic, Reasoning and Coding Capabilities
  2. OpenAIVideoProprietary

    Sora 2

    Sora 2 is OpenAI's flagship video and audio generation model, which OpenAI says is more physically accurate and controllable than prior systems and adds synchronized dialogue and sound effects. It launched with a new Sora social app.

    Modalities
    text, image → video, audio
    At launch
    Consumer app
    • Initial invite-based rollout in the U.S. and Canada via the Sora iOS app and sora.com
    • ChatGPT Pro users could use the higher-quality Sora 2 Pro model on sora.com
    • API models sora-2 and sora-2-pro released at DevDay on October 6, 2025
    Source: Sora 2 is here | OpenAI
  3. AnthropicLLMProprietary

    Claude Sonnet 4.5

    Claude Sonnet 4.5 is an Anthropic model aimed at coding, building complex agents, and computer use, with gains in reasoning and math. Anthropic released the Claude Agent SDK and API context editing and memory tools alongside it.

    Modalities
    text, image → text
    Context
    200K context
    At launch
    API, Consumer app, Cloud platforms
    • API model ID claude-sonnet-4-5 (snapshot claude-sonnet-4-5-20250929)
    • 200K-token context window and 64K max output tokens
    • Deprecated September 30, 2026, with Claude API retirement set for November 30, 2026
    Source: Introducing Claude Sonnet 4.5
  4. DeepSeekLLMOpen weights

    DeepSeek-V3.2-Exp

    DeepSeek-V3.2-Exp is an experimental model built on V3.1-Terminus that debuts DeepSeek Sparse Attention for more efficient long-context training and inference. DeepSeek says it performs on par with V3.1-Terminus.

    Modalities
    text → text
    License
    MIT
    At launch
    API, Consumer app, Open weights
    • Live on the DeepSeek app, web, and API at launch
    • V3.1-Terminus stayed available on a temporary API for comparison until October 15, 2025
    • Released with GPU kernels in TileLang and CUDA and a technical report
    Source: Introducing DeepSeek-V3.2-Exp
  5. TencentImageOpen weights

    HunyuanImage 3.0

    HunyuanImage 3.0 is an open native multimodal image generation model that Tencent describes as the largest open-source image generation mixture-of-experts model at release.

    Modalities
    text → image
    Parameters
    MoE: 80B total, 13B active
    License
    Tencent Hunyuan Community License Agreement
    At launch
    Open weights
    • 64 experts with 80B total parameters and 13B activated per token
    • Technical report published the same day
    • vLLM-accelerated inference added October 30, 2025
    Source: GitHub - Tencent-Hunyuan/HunyuanImage-3.0: HunyuanImage-3.0: A Powerful Native Multimodal Model for Image Generation
  6. TencentImageOpen weights

    Hunyuan3D-Omni

    Hunyuan3D-Omni is an open framework for controllable 3D asset generation that inherits the Hunyuan3D 2.1 structure and adds a unified control encoder for point cloud, voxel, skeleton, and bounding box conditions.

    Modalities
    image → 3d
    Parameters
    3.3B
    License
    Tencent Hunyuan 3D Omni Community License Agreement
    At launch
    Open weights
    • 3.3B-parameter image-to-shape model with multi-modal control
    • Requires about 10 GB of VRAM for generation
    Source: License.txt · tencent/Hunyuan3D-Omni at main
  7. GoogleMultimodalProprietary

    Gemini Robotics-ER 1.5

    Gemini Robotics-ER 1.5 is an embodied reasoning model that acts as a high-level planner for robots, and the first Gemini Robotics model made broadly available to developers.

    Modalities
    text, image, audio, video → text
    Context
    1M context
    At launch
    API
    • Preview in Google AI Studio and the Gemini API
    • A thinking model with a flexible thinking budget to trade latency for accuracy
    • Replaced by Gemini Robotics-ER 1.6 and shut down on Apr 30, 2026
    Source: Building the Next Generation of Physical Agents with Gemini Robotics-ER 1.5
  8. Liquid AISmall modelOpen weights

    Liquid Nanos (LFM2-Extract, LFM2-350M-ENJP-MT, LFM2-1.2B-RAG, LFM2-1.2B-Tool, LFM2-350M-Math)

    A first set of experimental task-specific LFM2 models for on-device agentic building blocks: data extraction, English-Japanese translation, retrieval-augmented question answering, tool calling, and math reasoning, released as open-weight 350M and 1.2B checkpoints.

    Modalities
    text → text
    Parameters
    350M and 1.2B checkpoints
    License
    LFM Open License v1.0
    At launch
    Open weights
    • Models: LFM2-350M-Extract, LFM2-1.2B-Extract, LFM2-350M-ENJP-MT, LFM2-1.2B-RAG, LFM2-1.2B-Tool, LFM2-350M-Math
    • Extract models turn unstructured documents into JSON, XML, or YAML
    • Available on Hugging Face and Liquid's LEAP platform for iOS, Android, and laptops
    Source: Introducing Liquid Nanos: frontier-grade performance on everyday devices
  9. MetaCodingOpen weights

    Code World Model (CWM)

    Code World Model is a 32B dense, decoder-only open-weights LLM from Meta FAIR, released to support research on code generation with world models.

    Modalities
    text → text
    Context
    128K context
    Parameters
    32B
    License
    FAIR Non-Commercial Research License
    At launch
    Open weights
    • Dense decoder-only model trained with up to 131k tokens of context
    • Checkpoints released after mid-training, SFT and RL
    • Intended for non-commercial research, not as a chat assistant
    Source: CWM: An Open-Weights LLM for Research on Code Generation with World Models
  10. Alibaba QwenLLMProprietary

    Qwen3-Max

    Qwen3-Max is Alibaba's largest Qwen model at launch, with more than 1 trillion parameters pre-trained on 36T tokens. The official release followed a preview and improved coding and agent capabilities.

    Modalities
    text → text
    Parameters
    MoE: over 1T total
    At launch
    API, Consumer app
    • Qwen3-Max-Instruct available via the Alibaba Cloud API and Qwen Chat
    • Trained with a 1M-token context using the ChunkFlow strategy
    • A Thinking variant was still in training at launch
    Source: Qwen3-Max: Just Scale it
  11. Liquid AISmall modelOpen weights

    LFM2-2.6B

    The largest dense model in the LFM2 series at its release, a 2.6B-parameter hybrid convolution and grouped query attention model trained on 10 trillion tokens and tuned for English and Japanese.

    Modalities
    text → text
    Context
    32K context
    Parameters
    2.6B
    License
    LFM Open License v1.0
    At launch
    Open weights
    • Also supports French, Spanish, German, Italian, Portuguese, Arabic, Chinese, and Korean
    • 32,768-token context window per model card
    • Released on Hugging Face under the LFM Open License
    Source: Introducing LFM2-2.6B: Redefining Efficiency in Language Models
  12. Alibaba QwenImageOpen weights

    Qwen-Image-Edit-2509

    Qwen-Image-Edit-2509 is the September 2025 monthly iteration of the open-weights Qwen-Image-Edit model, adding support for multi-image input and improving consistency in edits compared with the August release.

    Modalities
    text, image → image
    Parameters
    20B
    License
    Apache-2.0
    At launch
    Consumer app, Open weights
    • Multi-image editing support
    • Available through the Image Editing feature on Qwen Chat
    • Weights under the Apache 2.0 license
    Source: Qwen-Image-Edit-2509: Multi-Image Support, Improved Consistency
  13. Alibaba QwenSpeechProprietary

    Qwen3-LiveTranslate-Flash

    Qwen3-LiveTranslate-Flash is Qwen's real-time multilingual audio and video interpretation model built on Qwen3-Omni. It supports offline and live translation in 18 languages and uses visual context to improve translations.

    Modalities
    audio, video → text, audio
    At launch
    API
    • First Qwen interpretation model with visual context augmentation
    • Trained on millions of hours of multimodal data
    • Later followed by Qwen3.5-LiveTranslate-Flash and Qwen3.8-LiveTranslate
    Source: Qwen3-LiveTranslate: Real-Time Multimodal Interpretation
  14. Alibaba QwenMultimodalOpen weights

    Qwen3-VL

    Qwen3-VL is Qwen's vision-language series, opened with Qwen3-VL-235B-A22B in Instruct and Thinking versions. Qwen says its text performance matches the flagship Qwen3-235B-A22B-2507 language model.

    Modalities
    text, image, video → text
    Context
    250K context
    Parameters
    MoE: 235B total, 22B active (flagship)
    License
    Apache-2.0
    At launch
    API, Consumer app, Open weights
    • Native 256K-token context, expandable to 1M tokens
    • OCR expanded to 32 languages from 10
    • 30B-A3B followed October 4, 4B and 8B on October 15, and 2B and 32B on October 21, 2025
    Source: Qwen3-VL: Sharper Vision, Deeper Thought, Broader Action
  15. Alibaba QwenLLMOpen weights

    Qwen3Guard

    Qwen3Guard is the first safety guardrail model series in the Qwen family, built on Qwen3 and fine-tuned for safety classification. It includes Qwen3Guard-Gen and Qwen3Guard-Stream, which detects unsafe content in real time during generation.

    Modalities
    text → text
    Parameters
    0.6B, 4B, 8B
    License
    Apache-2.0
    At launch
    Open weights
    • Both variants come in 0.6B, 4B, and 8B sizes
    • Stream variant enables token-level streaming safety detection
    • Weights under the Apache 2.0 license
    Source: Qwen3Guard: Real-time Safety for Your Token Stream
  16. DeepSeekLLMOpen weights

    DeepSeek-V3.1-Terminus

    DeepSeek-V3.1-Terminus is an update to DeepSeek-V3.1 that addresses user feedback on language consistency and strengthens its code agent and search agent performance. It replaced V3.1 across the app, web, and API.

    Modalities
    text → text
    License
    MIT
    At launch
    API, Consumer app, Open weights
    • Fewer Chinese and English mix-ups and fewer abnormal characters
    • Upgraded both deepseek-chat (non-thinking) and deepseek-reasoner (thinking)
    • Weights on Hugging Face
    Source: DeepSeek-V3.1-Terminus
  17. MeituanReasoningOpen weights

    LongCat-Flash-Thinking

    LongCat-Flash-Thinking is Meituan's large reasoning model built on the LongCat-Flash mixture-of-experts architecture, released and open-sourced the same day with chat and API access.

    Modalities
    text → text
    Parameters
    MoE: 560B total, about 27B active
    License
    MIT
    At launch
    API, Consumer app, Open weights
    • 560B total parameters, sharing the LongCat-Flash architecture
    • Available free in LongCat Chat and via API as model=LongCat-Flash-Thinking
    • Weights on Hugging Face and GitHub under MIT
    Source: Change Log | LongCat API Docs
  18. BaiduMultimodalOpen weights

    Qianfan-VL

    Qianfan-VL is a Baidu AI Cloud family of vision-language models in 3B, 8B, and 70B sizes, continuously pre-trained for enterprise OCR and document understanding, with chain-of-thought reasoning in the 8B and 70B versions.

    Modalities
    image, text → text
    Context
    32K context
    Parameters
    3B, 8B, 70B
    License
    MIT
    At launch
    Open weights
    • 3B model built on Qwen2.5-3B; 8B and 70B models based on the Llama 3.1 architecture
    • InternViT-based vision encoder with dynamic patching up to 4K resolution
    • 32K context across all three sizes
    Source: GitHub - baidubce/Qianfan-VL: Qianfan-VL: Domain-Enhanced Universal Vision-Language Models
  19. Alibaba QwenMultimodalOpen weights

    Qwen3-Omni

    Qwen3-Omni is Qwen's natively end-to-end multilingual omni-modal model, processing text, images, audio, and video and responding in real time with text and natural speech. It was released as Instruct, Thinking, and Captioner models.

    Modalities
    text, image, audio, video → text, audio
    Parameters
    MoE: 30B total, 3B active
    License
    Apache-2.0
    At launch
    API, Open weights
    • Text interaction in 119 languages, speech understanding in 19, and speech generation in 10
    • Qwen3-Omni-30B-A3B-Captioner is an open audio captioning model
    • Weights under the Apache 2.0 license
    Source: Qwen3-Omni: Natively Omni-Modal Foundation Models!
  20. Alibaba QwenSpeechProprietary

    Qwen3-TTS-Flash

    Qwen3-TTS-Flash is Qwen's flagship text-to-speech model on the Qwen API, supporting multi-timbre, multilingual, and multi-dialect speech synthesis. It aims to produce natural and expressive speech.

    Modalities
    text → audio
    At launch
    API
    • 17 voices, each supporting 10 languages
    • Dialects including Hokkien, Wu, Cantonese, and Sichuanese
    • Available through the Qwen API on Alibaba Cloud Model Studio
    Source: Qwen3-TTS-Flash: Multi-timbre & Multi-lingual & Multi-dialect Speech Synthesis
  21. xAILLMProprietary

    Grok 4 Fast

    Grok 4 Fast is xAI's cost-efficient model built on learnings from Grok 4, with a unified architecture in which the same weights handle reasoning and non-reasoning modes, plus web and X search.

    Modalities
    text → text
    Context
    2M context
    At launch
    Consumer app, API
    • API model IDs: grok-4-fast-reasoning and grok-4-fast-non-reasoning, each with 2M-token context
    • Available to all grok.com and app users, including free users
    • Also offered on OpenRouter and Vercel AI Gateway
    Source: Grok 4 Fast | SpaceXSI
  22. KuaishouVideoProprietary

    Kling 2.5 Turbo

    Kling 2.5 Turbo is an upgraded Kling video model with improved text-to-video and image-to-video quality, including better prompt adherence, motion, style consistency, and aesthetics.

    Modalities
    text, image → video
    At launch
    Consumer app
    • Improved handling of multi-step instructions and causal relationships in prompts
    • Larger-amplitude motion and camera movement with more realistic physics
    • Kuaishou's press release dates the launch to September 23, 2025
    Source: Kling AI Release Notes & Version History | Kling AI
  23. XiaomiSpeechOpen weights

    MiMo-Audio

    MiMo-Audio is Xiaomi's open-source audio language model, pretrained on more than 100 million hours of audio and showing few-shot learning on audio tasks such as voice conversion, style transfer, and speech continuation.

    Modalities
    text, audio → text, audio
    Parameters
    7B
    License
    MIT
    At launch
    Open weights
    • Released as MiMo-Audio-7B-Base and MiMo-Audio-7B-Instruct
    • Includes MiMo-Audio-Tokenizer, a 1.2B-parameter 25 Hz tokenizer
    • Instruct model adds thinking to audio understanding and generation
    • Technical report dated September 19, 2025 on the MiMo site
    Source: XiaomiMiMo/MiMo-Audio-7B-Instruct
  24. Alibaba QwenVideoOpen weights

    Wan2.2-Animate-14B

    Wan2.2-Animate-14B is a unified model for character animation and replacement that replicates full-body movement and facial expression from a reference video.

    Modalities
    text, image, video → video
    Parameters
    14B
    License
    Apache-2.0
    At launch
    Open weights
    • Model weights and inference code released
    • Diffusers integration added November 13, 2025
    • Released under the Apache 2.0 license
    Source: Wan2.2 (GitHub)
  25. Luma AIVideoProprietary

    Ray3

    Video model that Luma describes as the first reasoning video model, able to evaluate and refine its own outputs, and the first to generate 10, 12 and 16-bit HDR video in ACES2065-1 EXR format.

    Modalities
    text, image → video
    At launch
    Consumer app
    • Native HDR EXR generation
    • Draft Mode for faster exploration
    • Launched in Dream Machine with Adobe Firefly as first outside partner
    Source: Luma AI launches Ray3
  26. IBMMultimodalOpen weights

    Granite-Docling 258M

    Compact document conversion vision-language model built on the Idefics3 architecture with a SigLIP2 vision encoder and a Granite 165M language model.

    Modalities
    text, image → text
    Parameters
    258M
    License
    Apache-2.0
    At launch
    Open weights
    • Builds on the Idefics3 architecture
    • Replaces the vision encoder with siglip2-base-patch16-512
    • Released under Apache 2.0
    Source: ibm-granite/granite-docling-258M · Hugging Face
  27. Mistral AIReasoningProprietary

    Magistral Medium 1.2

    September 2025 update of Magistral Medium that turns it into a multimodal reasoning model and extends its context window to 128K tokens.

    Modalities
    text, image → text
    Context
    125K context
    At launch
    API
    • API model ID magistral-medium-2509
    • Mistral's model card lists September 18, 2025
    • Deprecated May 22, 2026 in favor of Mistral Medium 3.5
    Source: Changelog | Mistral Docs
  28. Mistral AIReasoningOpen weights

    Magistral Small 1.2

    September 2025 update of the open Magistral Small reasoning model that adds image input and raises the context window from 40K to 128K tokens.

    Modalities
    text, image → text
    Context
    125K context
    Parameters
    24B
    License
    Apache-2.0
    At launch
    Open weights, API
    • API model ID magistral-small-2509
    • Mistral's model card lists September 18, 2025
    • Deprecated April 30, 2026 in favor of Mistral Small 4
    Source: Changelog | Mistral Docs
  29. RekaSpeechProprietary

    Reka Speech

    An 850M-parameter multilingual speech model for transcription and translation with timestamps, aimed at high-volume offline audio jobs and using a QK-cache alignment method on a modified vLLM for fast timestamping.

    Modalities
    audio → text
    Parameters
    850M (550M decoder, 300M audio encoder)
    At launch
    Enterprise
    • Understands English, Japanese, Korean, Chinese, French, Spanish, German, Italian, Portuguese and Arabic
    • Translates from English to each of the other supported languages and vice versa
    • Available to customers by contacting Reka
    Source: Reka Speech: High Throughput Speech Transcription and Translation Model with Timestamps
  30. OpenAICodingProprietary

    GPT-5-Codex

    GPT-5-Codex is a version of GPT-5 further optimized for agentic coding in Codex, trained on real-world software engineering work for both quick interactive sessions and long independent tasks, including code review.

    Modalities
    text, image → text
    Context
    400K context
    At launch
    Consumer app, Enterprise
    • Default for Codex cloud tasks and code review; selectable in Codex CLI and the IDE extension
    • Included with ChatGPT Plus, Pro, Business, Edu, and Enterprise plans
    • Added to the Responses API as gpt-5-codex on September 23, 2025
    • 400,000-token context window per the API model page
    Source: Introducing upgrades to Codex | OpenAI
  31. GoogleSmall modelOpen weights

    VaultGemma

    VaultGemma is a 1B-parameter Gemma 2-based model that Google Research trained from scratch with differential privacy. Google describes it as the largest open model trained this way, with a formal sequence-level privacy guarantee.

    Modalities
    text → text
    Context
    1K context
    Parameters
    1B
    License
    Gemma Terms of Use
    At launch
    Open weights
    • Sequence-level DP guarantee of epsilon at most 2.0 and delta at most 1.1e-10, over 1,024-token sequences
    • Pre-trained checkpoint only
    • Weights on Hugging Face and Kaggle with a technical report
    Source: VaultGemma: The world's most capable differentially private LLM
  32. Alibaba QwenLLMOpen weights

    Qwen3-Next

    Qwen3-Next is a new Qwen architecture for training and inference efficiency, launched with Qwen3-Next-80B-A3B in Instruct and Thinking versions. It combines hybrid attention with an ultra-sparse mixture-of-experts design.

    Modalities
    text → text
    Context
    256K context
    Parameters
    MoE: 80B total, 3B active
    License
    Apache-2.0
    At launch
    API, Open weights
    • Base model matches dense Qwen3-32B at under 10% of its training cost, per Qwen
    • More than 10 times higher throughput at contexts over 32K tokens
    • Native 262,144-token context, extensible to 1,010,000 tokens
    Source: Qwen3-Next: Towards Ultimate Training & Inference Efficiency
  33. Stability AIMusicProprietary

    Stable Audio 2.5

    Enterprise-focused audio model with improved musical composition that generates tracks up to three minutes long in under two seconds on a GPU and supports audio inpainting.

    Modalities
    text, audio → audio
    At launch
    API, Consumer app, Cloud platforms, Enterprise
    • Post-trained with the Adversarial Relativistic-Contrastive method
    • Trained on a fully licensed dataset
    • Available via the Stability API, fal, Replicate, ComfyUI and on-premises licensing
    Source: Stability AI Introduces Stable Audio 2.5, the First Audio Model Built for Enterprise Sound Production at Scale
  34. BaiduReasoningProprietary

    ERNIE X1.1

    ERNIE X1.1 upgrades Baidu's ERNIE X1 deep-thinking model, using an iterative hybrid reinforcement learning framework to improve factuality, instruction following, and agentic task performance.

    Modalities
    text → text
    At launch
    Consumer app, API
    • Announced at WAVE SUMMIT 2025
    • Available on the ERNIE Bot website, the Wenxiaoyan app, and Baidu AI Cloud Qianfan
    • Like ERNIE X1, trained from ERNIE 4.5
    Source: 文心大模型X1.1,正式发布!
  35. BaiduReasoningOpen weights

    ERNIE-4.5-21B-A3B-Thinking

    ERNIE-4.5-21B-A3B-Thinking is an open-weights deep-thinking model trained from ERNIE-4.5-21B-A3B for reasoning, math, code, and tool use, released in Transformer-style weights for PyTorch and PaddlePaddle tooling.

    Modalities
    text → text
    Context
    128K context
    Parameters
    MoE: 21B total, 3B active
    License
    Apache-2.0
    At launch
    Open weights
    • Open-sourced at WAVE SUMMIT 2025 on September 9, 2025
    • Enhanced 128K long-context understanding
    • Runs with vLLM, transformers, and FastDeploy
    Source: baidu/ERNIE-4.5-21B-A3B-Thinking · Hugging Face
  36. ByteDance SeedImageProprietary

    Seedream 4.0

    Seedream 4.0 is ByteDance Seed's image creation model that unifies text-to-image generation and general-purpose editing in one architecture, adding commonsense reasoning, multi-image reference, and output up to 4K resolution.

    Modalities
    text, image → image
    At launch
    Consumer app, API
    • Succeeds both Seedream 3.0 and SeedEdit 3.0
    • Maximum resolution raised from 2K to 4K
    • Supports in-context reasoning and consistent image sequences
    • Available on Jimeng, Doubao, and Volcano Ark
    Source: Seedream 4.0 Officially Released: Beyond Drawing, Into Imagination
  37. IBMRerankerOpen weights

    Granite Embedding Reranker R2

    Cross-encoder reranking model from IBM's Granite Embeddings R2 collection that computes relevance scores between queries and documents by encoding them jointly for precise reranking.

    Modalities
    text → score
    Parameters
    149M
    License
    Apache-2.0
    At launch
    Open weights
    • Part of the Granite Embeddings R2 collection
    • Cross-encoder architecture
    • Released under Apache 2.0
    Source: ibm-granite/granite-embedding-reranker-english-r2 · Hugging Face
  38. TencentImageOpen weights

    HunyuanImage 2.1

    HunyuanImage 2.1 is an open 17B text-to-image diffusion model that generates 2K (2048x2048) images, using a multimodal LLM and a character-aware multilingual encoder for image-text alignment and text rendering.

    Modalities
    text → image
    Parameters
    17B
    License
    Tencent Hunyuan Community License Agreement
    At launch
    Open weights
    • Native 2K resolution output
    • FP8 quantized models released September 12, 2025
    Source: GitHub - Tencent-Hunyuan/HunyuanImage-2.1: HunyuanImage-2.1: An Efficient Diffusion Model for High-Resolution (2K) Text-to-Image Generation
  39. Alibaba QwenSpeechProprietary

    Qwen3-ASR-Flash

    Qwen3-ASR-Flash is a speech recognition service built on Qwen3-Omni and tens of millions of hours of speech data. It supports 11 languages and multiple accents and accepts text context to customize results.

    Modalities
    audio → text
    At launch
    API
    • Handles singing voice recognition
    • Users can prompt with context in any format for customized transcription
    • Offered as an API service
    Source: Qwen3 ASR: Hear clearly, transcribe smartly.
  40. Moonshot AILLMOpen weights

    Kimi K2 Instruct 0905

    Kimi-K2-Instruct-0905 is an updated Kimi K2 instruct model that extends the context window from 128K to 256K tokens. Moonshot also improved its agentic coding and front-end programming abilities.

    Modalities
    text → text
    Context
    250K context
    Parameters
    MoE: 1T total, 32B active
    License
    Modified MIT License
    At launch
    API, Open weights
    • Context length doubled to 256K tokens
    • Offered on the Kimi Open Platform as kimi-k2-0905-preview
    • Weights and code under the Modified MIT License
    Source: moonshotai/Kimi-K2-Instruct-0905 (Hugging Face)
  41. GoogleEmbeddingOpen weights

    EmbeddingGemma

    EmbeddingGemma is a 308M-parameter open text embedding model based on the Gemma 3 architecture and built for on-device use. It is trained on more than 100 languages and runs in under 200MB of RAM when quantized.

    Modalities
    text → embedding
    Context
    2K context
    Parameters
    308M
    License
    Gemma Terms of Use
    At launch
    Open weights, Cloud platforms
    • 2K-token context window
    • 768-dimension output, truncatable to 512, 256 or 128 via Matryoshka representation
    • Weights on Hugging Face, Kaggle and Vertex AI
    Source: Introducing EmbeddingGemma: The Best-in-Class Open Model for On-Device Embeddings
  42. TencentVideoOpen weights

    HunyuanWorld-Voyager

    HunyuanWorld-Voyager is an open video diffusion model that generates world-consistent RGB-D video and 3D point-cloud sequences from a single image along a user-defined camera path, supporting world exploration and 3D reconstruction.

    Modalities
    text, image → video
    License
    Tencent HunyuanWorld-Voyager Community License Agreement
    At launch
    Open weights
    • Generates aligned depth and RGB video for direct 3D reconstruction
    • Trained with a data engine that curated over 100,000 video clips
    Source: GitHub - Tencent-Hunyuan/HunyuanWorld-Voyager: Voyager is an interactive RGBD video generation model conditioned on camera input, and supports real-time 3D reconstruction.
  43. TencentLLMOpen weights

    Hunyuan-MT-7B and Hunyuan-MT-Chimera-7B

    Hunyuan-MT is Tencent's open translation release, pairing the Hunyuan-MT-7B translation model with Hunyuan-MT-Chimera-7B, an ensemble model that merges multiple translation outputs into a higher-quality result.

    Modalities
    text → text
    Parameters
    7B
    License
    Tencent Hunyuan Community License Agreement
    At launch
    Open weights
    • Supports mutual translation among 33 languages, including five ethnic minority languages in China
    • Tencent describes Chimera as the industry's first open-source translation ensemble model
    • FP8 versions also published
    Source: GitHub - Tencent-Hunyuan/Hunyuan-MT

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 September 2025." The AI Engineering, updated Sep 30, 2025. https://theaiengineering.com/releases/2025/09

HTML link

<a href="https://theaiengineering.com/releases/2025/09">AI models released in September 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