AI models released in December 2025

53 releases from 24 labs in December 2025, including Qwen-Image-2512, HY-Motion 1.0 and HY-MT1.5. Every entry is dated to the lab's official announcement and links to it.

Releases
53
Labs
24
Open weights
33

December 2025

53 releases
  1. Alibaba QwenImageOpen weights

    Qwen-Image-2512

    Qwen-Image-2512 is the December 2025 update of Qwen-Image's text-to-image foundation model. Qwen says it reduces the AI-generated look, improves realism for human subjects, and renders finer natural detail.

    Modalities
    text → image
    Parameters
    20B
    License
    Apache-2.0
    At launch
    Consumer app, Open weights
    • Weights on Hugging Face and ModelScope on release day
    • Available to try on Qwen Chat
    • Released under the Apache 2.0 license
    Source: Qwen-Image-2512: Finer Details, Greater Realism
  2. TencentImageOpen weights

    HY-Motion 1.0

    HY-Motion 1.0 is an open series of text-to-3D human motion generation models based on a Diffusion Transformer and flow matching, producing skeleton-based 3D character animations from text prompts.

    Modalities
    text → 3d
    Parameters
    1.0B
    License
    Tencent HY-Motion 1.0 Community License Agreement
    At launch
    Open weights
    • Standard model with 1.0B parameters and a Lite model with 0.46B
    • Tencent describes it as the first DiT-based text-to-motion model scaled to a billion parameters
    Source: GitHub - Tencent-Hunyuan/HY-Motion-1.0: HY-Motion model for 3D human motion or 3D character animation generation.
  3. TencentLLMOpen weights

    HY-MT1.5

    HY-MT1.5 is Tencent's second open translation model release, with 1.8B and 7B models supporting mutual translation across 33 languages and five ethnic and dialect variations. The 1.8B model targets edge devices and real-time translation.

    Modalities
    text → text
    Parameters
    1.8B, 7B
    License
    Tencent HY Community License Agreement
    At launch
    Open weights
    • HY-MT1.5-7B is optimized for annotated and mixed-language text over the September 2025 release
    • FP8, GPTQ-Int4, and GGUF versions published
    Source: GitHub - Tencent-Hunyuan/Hy-MT
  4. StepFunImageOpen weights

    NextStep-1.1

    NextStep-1.1 is an updated version of StepFun's autoregressive text-to-image model that improves output quality through extended training and a flow-based reinforcement learning post-training stage.

    Modalities
    text → image
    License
    Apache-2.0
    At launch
    Open weights
    • Builds on NextStep-1's autoregressive generation with continuous image tokens
    • Adds flow-based reinforcement learning post-training
    • Checkpoints hosted on Hugging Face under Apache-2.0
    Source: GitHub - stepfun-ai/NextStep-1
  5. MiniMaxCodingOpen weights

    MiniMax M2.1

    MiniMax M2.1 updates M2 for real-world complex tasks, with stronger programming in Rust, Java, Go, C++, Kotlin, Objective-C, TypeScript, and JavaScript, plus improved web and mobile app development and office scenarios.

    Modalities
    text → text
    License
    Modified-MIT
    At launch
    API, Open weights
    • Released on the API as MiniMax-M2.1 and M2.1-highspeed
    • Weights open-sourced on Hugging Face under a modified MIT license
    • Powers the MiniMax Agent product
    Source: MiniMax M2.1: Significantly Enhanced Multi-Language Programming, Built for Real-World Complex Tasks
  6. Alibaba QwenImageOpen weights

    Qwen-Image-Edit-2511

    Qwen-Image-Edit-2511 is an enhanced version of Qwen-Image-Edit-2509 with better consistency. It mitigates image drift and improves character consistency in edits.

    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-2511: Improve Consistency
  7. Z.aiLLMOpen weights

    GLM-4.7

    GLM-4.7 is Z.ai's foundation model with improvements in coding, reasoning, and agentic capabilities over GLM-4.6. It supports thinking before acting in agent frameworks such as Claude Code, Kilo Code, Cline, and Roo Code.

    Modalities
    text → text
    Context
    200K context
    Parameters
    MoE: 355B total, 32B active
    License
    MIT
    At launch
    API, Open weights
    • Gains in multilingual agentic coding and terminal-based tasks
    • Cleaner webpage and slide generation
    • Weights on Hugging Face and ModelScope under the MIT license
    Source: GLM-4.7: Advancing the Coding Capability
  8. MeituanLLMProprietary

    LongCat-Flash-Chat (December 2025 update)

    Meituan upgraded the LongCat-Flash-Chat model on its API with stronger programming performance, a doubled context length, broader multilingual support, and more robust agent behavior, without changing the model name.

    Modalities
    text → text
    Context
    250K context
    At launch
    API, Consumer app
    • Context length doubled from 128K to 256K tokens
    • High-quality support for nine languages, including Spanish, French, and Arabic
    • Model name and API invocation unchanged
    Source: Change Log | LongCat API Docs
  9. Alibaba QwenImageOpen weights

    Qwen-Image-Layered

    Qwen-Image-Layered is a Qwen model that decomposes an image into multiple RGBA layers, so each layer can be edited independently. The layered output supports resizing, repositioning, and recoloring without affecting other content.

    Modalities
    text, image → image
    License
    Apache-2.0
    At launch
    Open weights
    • Weights on Hugging Face and ModelScope
    • Research paper published December 18, 2025
    • Released under the Apache 2.0 license
    Source: Qwen-Image-Layered: Layered Decomposition for Inherent Editablity
  10. GoogleSmall modelOpen weights

    FunctionGemma

    FunctionGemma is a version of Gemma 3 270M fine-tuned for function calling, intended as a base for fast, private, on-device agents. It can handle tasks offline or route complex requests to larger models.

    Modalities
    text → text
    Context
    32K context
    Parameters
    270M
    License
    Gemma Terms of Use
    At launch
    Open weights, Cloud platforms
    • Small enough to run on NVIDIA Jetson Nano and mobile phones
    • 32K-token input context per the model card
    • Weights on Hugging Face and Kaggle
    Source: FunctionGemma: Bringing bespoke function calling to the edge
  11. OpenAICodingProprietary

    GPT-5.2-Codex

    GPT-5.2-Codex is a version of GPT-5.2 further optimized for agentic coding in Codex, with better long-horizon work through context compaction, stronger handling of large refactors and migrations, and stronger cybersecurity capabilities.

    Modalities
    text, image → text
    Context
    400K context
    At launch
    Consumer app
    • Released in all Codex surfaces for paid ChatGPT users; CLI and IDE extension default to gpt-5.2-codex
    • API model ID gpt-5.2-codex reached the Responses API on January 14, 2026
    • Improved performance in Windows environments
    • 400,000-token context window per the API model page
    Source: Introducing GPT-5.2-Codex | OpenAI
  12. Luma AIVideoProprietary

    Ray3 Modify

    Ray3-based video-to-video model for editing and reimagining footage, adding start and end frame keyframe control and character references while preserving the original performance.

    Modalities
    text, image, video → video
    At launch
    Consumer app
    • First start and end frame control in Luma's video-to-video workflow
    • Character Reference for consistent subjects
    • Available in Dream Machine
    Source: Introducing Ray3 Modify
  13. ByteDance SeedReasoningProprietary

    Seed1.8

    Seed1.8 is a generalized agentic model that combines search, coding, and GUI agent capabilities with native vision, accepting text and image input and offering three thinking modes to balance latency against depth.

    Modalities
    text, image → text
    At launch
    API
    • Integrates search, code, and GUI agent abilities in one model
    • Three thinking modes adjust to task complexity
    • Reduces the tokens needed to encode images and video
    • Available on Volcano Engine as Doubao 1.8
    Source: Official Release of Seed1.8: A Generalized Agentic Model
  14. GoogleLLMOpen weights

    T5Gemma 2

    T5Gemma 2 is the second generation of Google's encoder-decoder Gemma models, now based on Gemma 3. It adds image understanding, a 128K-token context window and support for over 140 languages.

    Modalities
    text, image → text
    Context
    128K context
    Parameters
    270M-270M, 1B-1B, 4B-4B
    License
    Gemma Terms of Use
    At launch
    Open weights, Cloud platforms
    • Sizes total about 370M, 1.7B and 7B parameters, excluding the vision encoder
    • Handles visual question answering and multimodal reasoning
    • Only pre-trained checkpoints released, no instruction-tuned versions
    Source: T5Gemma 2: The next generation of encoder-decoder models
  15. GoogleLLMProprietary

    Gemini 3 Flash

    Gemini 3 Flash brings Gemini 3 frontier intelligence to a fast, lower-cost Flash model with upgraded visual and spatial reasoning and agentic coding, rolling out globally in the Gemini app and AI Mode in Search.

    Modalities
    text, image, audio, video → text
    Context
    1M context
    At launch
    API, Cloud platforms, Consumer app, Enterprise
    • Preview model gemini-3-flash-preview in the Gemini API, Google Antigravity, Gemini CLI, Android Studio and Vertex AI
    • Also available in Gemini Enterprise
    • Input token limit of 1,048,576 per the Gemini API model page
    Source: Introducing Gemini 3 Flash: Benchmarks, global availability
  16. xAISpeechProprietary

    Grok Voice Agent API

    The Grok Voice Agent API opened xAI's in-house speech-to-speech stack, which powers Grok Voice in its apps and Tesla vehicles, to developers building multilingual voice agents that call tools and search real-time data.

    Modalities
    audio → audio
    At launch
    API
    • Voice activity detection, tokenizer, and audio models trained in-house from scratch
    • Average time-to-first-audio under 1 second, per xAI
    • Compatible with the OpenAI Realtime API specification; LiveKit plugin available
    • Voices include Ara, Eve, and Leo
    Source: Grok Voice Agent API | SpaceXSI
  17. TencentVideoOpen weights

    HY-World 1.5 (WorldPlay)

    HY-World 1.5 (WorldPlay) is an open streaming video diffusion model for real-time interactive world modeling with long-term geometric consistency, extending HunyuanWorld 1.0 beyond offline 3D world generation.

    Modalities
    text, image → video
    License
    Tencent HY-WorldPlay Community License Agreement
    At launch
    Open weights
    • WorldPlay-8B is based on HunyuanVideo
    • A lighter WorldPlay-5B model and training code followed on January 6, 2026
    Source: GitHub - Tencent-Hunyuan/HY-WorldPlay: HY-World 1.5: A Systematic Framework for Interactive World Modeling with Real-Time Latency and Geometric Consistency
  18. Mistral AIMultimodalProprietary

    Mistral OCR 3

    OCR model that extracts text and embedded images from forms, scans, tables and handwriting, returning markdown with HTML-based table reconstruction.

    Modalities
    text, image → text
    At launch
    API
    • Powers the Document AI playground in Mistral AI Studio
    • API model ID mistral-ocr-2512
    • Major upgrade over Mistral OCR 2 on handwriting and low-quality scans
    Source: Introducing Mistral OCR 3 | Mistral AI
  19. Black Forest LabsImageProprietary

    FLUX.2 [max]

    Highest-quality FLUX.2 model, adding grounding search that generates images from real-time web information, alongside the strongest editing consistency and prompt following in the FLUX.2 family.

    Modalities
    text, image → image
    At launch
    API, Consumer app
    • Grounding search for current events, weather and news imagery
    • Multi-reference editing with up to 10 input images
    • Available in the BFL Playground and API
    Source: Release Notes - Black Forest Labs
  20. OpenAIImageProprietary

    GPT Image 1.5

    GPT Image 1.5 is OpenAI's new flagship image generation model, powering a new version of ChatGPT Images and released in the API. OpenAI says it makes more precise edits that keep details intact and generates images up to 4x faster.

    Modalities
    text, image → image, text
    At launch
    Consumer app, API
    • API model ID gpt-image-1.5 (snapshot gpt-image-1.5-2025-12-16), plus chatgpt-image-latest
    • Rolled out to all ChatGPT users as the model behind the new ChatGPT Images
    • New Images creation space in ChatGPT; Business and Enterprise access came later
    Source: The new ChatGPT Images is here | OpenAI
  21. MeituanVideoOpen weights

    LongCat-Video-Avatar

    LongCat-Video-Avatar is a unified open model for expressive, highly dynamic audio-driven character video generation, built on Meituan's LongCat-Video foundation model and released with a technical report.

    Modalities
    text, image, audio → video
    License
    MIT
    At launch
    Open weights
    • Supports audio-text-to-video and audio-image-to-video generation
    • Handles single- and multi-speaker audio inputs
    • Weights released under MIT
    Source: GitHub - meituan-longcat/LongCat-Video
  22. XiaomiLLMOpen weights

    MiMo-V2-Flash

    MiMo-V2-Flash is Xiaomi's open-weight mixture-of-experts model for reasoning, coding, and agentic tasks, using hybrid sliding-window and global attention plus multi-token prediction for fast, efficient inference.

    Modalities
    text → text
    Context
    250K context
    Parameters
    MoE: 309B total, 15B active
    License
    MIT
    At launch
    API, Consumer app, Open weights
    • Hybrid attention at a 5:1 ratio with a 128-token sliding window
    • Hybrid thinking mode that can be switched on or off
    • Weights, inference code, and 3-layer MTP weights released under MIT
    • API updates on January 12 and February 4, 2026 improved coding and tool calling
    Source: MiMo-V2-Flash | Xiaomi
  23. Mistral AILLMProprietary

    Mistral Small Creative

    Experimental Labs model from Mistral trained on curated data for creative writing, narrative generation, roleplay, character-driven dialogue and general-purpose conversational agents.

    Modalities
    text → text
    Context
    32K context
    At launch
    API
    • API model ID labs-mistral-small-creative
    • Mistral's model card lists December 11, 2025
    • Deprecated March 31, 2026 in favor of Ministral 3 8B
    Source: Changelog | Mistral Docs
  24. MetaMultimodalOpen weights

    SAM Audio

    SAM Audio extends the Segment Anything approach to sound, a unified model that isolates specific sounds from audio mixtures using text, visual or time-span prompts.

    Modalities
    audio, text, video → audio
    Parameters
    500M to 3B
    License
    SAM License
    At launch
    Open weights
    • Prompts can be text, clicks on sounding objects in video, or marked time spans
    • Released alongside the PE-AV perception encoder
    • Audio prompts are not supported
    Source: Introducing SAM Audio: The First Unified Multimodal Model for Audio Separation
  25. ByteDance SeedVideoProprietary

    Seedance 1.5 pro

    Seedance 1.5 pro is ByteDance Seed's audio-visual generation model, adding joint audio and video synthesis to Seedance with precise lip and audio synchronization, multilingual and dialect speech, and stronger narrative coherence.

    Modalities
    text, image → video, audio
    At launch
    Consumer app, API
    • Generates synchronized audio and video together
    • Supports multiple languages and Chinese dialects in generated speech
    • Available in Jimeng AI and the Doubao app at launch
    • Volcano Engine API version supports first and last frame control
    Source: Sound and Vision, All in One Take | The Official Release of Seedance 1.5 pro
  26. Ai2LLMOpen weights

    Bolmo

    Family of byte-level language models created by byteifying Olmo 3 models, reusing their backbones while replacing subword tokenization with byte-level patching.

    Modalities
    text → text
    Parameters
    1B and 7B
    License
    Apache-2.0
    At launch
    Open weights
    • Uses an mLSTM local encoder and a boundary predictor
    • Built on the public Olmo 3 7B checkpoint
    • Released under Apache 2.0
    Source: Introducing Bolmo: Byteifying the next generation of language models | Ai2
  27. NVIDIAReasoningOpen weights

    Nemotron 3 Nano

    First model of the Nemotron 3 open family, a hybrid Mamba-Transformer mixture-of-experts model with a 1M-token context window for agentic workloads.

    Modalities
    text → text
    Context
    1M context
    Parameters
    MoE: 30B total, 3.5B active
    License
    NVIDIA Nemotron Open Model License
    At launch
    Open weights, Cloud platforms
    • Each MoE layer has 128 experts plus a shared expert
    • Super and Ultra models were announced for the first half of 2026
    • Available on Hugging Face and through inference providers
    Source: NVIDIA Debuts Nemotron 3 Family of Open Models | NVIDIA Newsroom
  28. Ai2ReasoningOpen weights

    Olmo 3.1

    Update to Olmo 3 with two new 32B checkpoints, Olmo 3.1 Think 32B and Olmo 3.1 Instruct 32B, plus refreshed RL Zero 7B models.

    Modalities
    text → text
    Parameters
    32B
    License
    Apache-2.0
    At launch
    Open weights
    • Think 32B came from 21 more days of RL training on 224 GPUs
    • Instruct 32B applies the Olmo 3 Instruct 7B recipe at larger scale
    • Released under Apache 2.0
    Source: Olmo 3: Charting a path through the model flow to lead open-source AI | Ai2
  29. Z.aiSpeechOpen weights

    GLM-TTS

    GLM-TTS is Z.ai's controllable, emotion-expressive zero-shot text-to-speech model, trained with multi-reward reinforcement learning. Z.ai launched it on its API alongside GLM-TTS-Clone for voice cloning from about 3 seconds of audio.

    Modalities
    text, audio → audio
    License
    MIT
    At launch
    API, Open weights
    • Two-stage generation architecture
    • Streaming and non-streaming interfaces
    • Weights and inference scripts open-sourced under the MIT license
    Source: GLM-TTS (GitHub)
  30. OpenAILLMProprietary

    GPT-5.2 (Instant, Thinking, and Pro)

    GPT-5.2 is the next model series in the GPT-5 line, released as Instant, Thinking, and Pro in ChatGPT and the API. OpenAI positions it for professional knowledge work such as spreadsheets, presentations, coding, and long-running agents.

    Modalities
    text, image → text
    At launch
    Consumer app, API, Enterprise
    • API IDs: gpt-5.2 (Thinking), gpt-5.2-chat-latest (Instant), and gpt-5.2-pro (Responses API)
    • Adds a fifth reasoning effort level, xhigh, for GPT-5.2 Thinking and Pro
    • Context: 400,000 tokens for gpt-5.2 and gpt-5.2-pro; 128,000 for gpt-5.2-chat-latest
    • ChatGPT rollout began with paid plans, including Business and Enterprise
    Source: Introducing GPT-5.2 | OpenAI
  31. Ai2MultimodalOpen weights

    Molmo 2

    Successor to Ai2's Molmo that extends its grounded vision capabilities to video and multi-image understanding, adding video pointing and tracking, released with open weights.

    Modalities
    text, image, video → text
    Parameters
    4B, 7B and 8B
    License
    Apache-2.0
    At launch
    Open weights
    • Molmo 2 8B and 4B are built on Qwen 3
    • Molmo 2-O 7B is built on Olmo for a fully open stack
    • Full training codebase released later
    Source: Molmo 2: State-of-the-art video understanding, pointing, and tracking | Ai2
  32. CohereRerankerProprietary

    Rerank 4

    Fourth-generation reranker with a 32K-token context window, four times the previous generation, offered in Pro and Fast variants for enterprise AI search.

    Modalities
    text → score
    Context
    32K context
    At launch
    API
    • Model IDs rerank-v4.0-pro and rerank-v4.0-fast
    • Covers 100+ languages
    • Key component of the North agent platform
    Source: Rerank 4: Cohere's Most Powerful Reranker Yet | Cohere
  33. Z.aiVideoOpen weights

    SCAIL-Preview

    SCAIL-Preview is the preview release of SCAIL, Z.ai's studio-grade character animation model that learns 3D-consistent pose representations in context. It animates a reference character from a driving pose sequence.

    Modalities
    text, image, video → video
    License
    MIT
    At launch
    Open weights
    • Inference code released December 8, 2025; weights followed December 11
    • Weights on Hugging Face and ModelScope
    • Released under the MIT license
    Source: SCAIL: Towards Studio-Grade Character Animation (GitHub)
  34. Z.aiSpeechOpen weights

    GLM-ASR-2512 and GLM-ASR-Nano-2512

    GLM-ASR-2512 is Z.ai's speech recognition model on its API, released with GLM-ASR-Nano-2512, a 1.5B open-weights version. They support Mandarin, Cantonese, other dialects, and dozens of languages.

    Modalities
    audio → text
    Parameters
    1.5B (Nano)
    License
    MIT
    At launch
    API, Open weights
    • Native custom dictionary support for domain terms
    • GLM-ASR-Nano-2512 is aimed at challenging acoustic environments
    • Nano weights on Hugging Face under the MIT license
    Source: Model and product release records (BigModel docs)
  35. Mistral AICodingOpen weights

    Devstral 2

    Next-generation 123B dense coding model for code agents, released under a modified MIT license together with the Mistral Vibe command-line agent.

    Modalities
    text → text
    Context
    250K context
    Parameters
    123B
    License
    Modified MIT License
    At launch
    Open weights, API
    • 256K-token context window
    • Deprecated May 22, 2026 in favor of Mistral Medium 3.5
    • API model ID devstral-2512
    Source: Introducing: Devstral 2 and Mistral Vibe CLI. | Mistral AI
  36. Mistral AICodingOpen weights

    Devstral Small 2

    24B coding model released alongside Devstral 2 under Apache 2.0, available through the API or deployable locally on consumer hardware.

    Modalities
    text → text
    Context
    250K context
    Parameters
    24B
    License
    Apache-2.0
    At launch
    Open weights, API
    • 256K-token context window
    • API model ID labs-devstral-small-2512
    • Supports on-premises deployment and fine-tuning
    Source: Introducing: Devstral 2 and Mistral Vibe CLI. | Mistral AI
  37. Nous ResearchReasoningOpen weights

    Nomos 1

    30B open-weights mathematics model released by Nous Research in partnership with Hillclimb, described by Nous as a small but capable mathematician for problem solving.

    Modalities
    text → text
    Parameters
    30B
    License
    Apache-2.0
    At launch
    Open weights
    • Focused on mathematical problem solving
    • Built in partnership with Hillclimb
    • Published on Hugging Face as nomos-1
    Source: Releases | Nous Research
  38. Z.aiMultimodalOpen weights

    AutoGLM-Phone-9B

    AutoGLM-Phone-9B is Z.ai's open phone agent model, released with the Open-AutoGLM framework for completing app tasks on a phone from natural language. Its architecture is identical to GLM-4.1V-9B-Thinking.

    Modalities
    text, image → text
    Parameters
    9B
    License
    MIT
    At launch
    API, Open weights
    • Released with a multilingual variant, AutoGLM-Phone-9B-Multilingual
    • Z.ai launched the AutoGLM-Phone framework on its API on December 11, 2025
    • Weights under the MIT license
    Source: Open-AutoGLM (GitHub)
  39. Z.aiMultimodalOpen weights

    GLM-4.6V

    GLM-4.6V is Z.ai's open-weights vision-language series, with GLM-4.6V for cloud use and GLM-4.6V-Flash for local, low-latency use. Z.ai says it integrates native function calling for the first time, passing images directly as tool parameters.

    Modalities
    text, image, video, file → text
    Context
    125K context
    Parameters
    106B (GLM-4.6V); 9B (GLM-4.6V-Flash)
    License
    MIT
    At launch
    API, Open weights
    • Visual context window scaled to 128K tokens in training
    • Generates interleaved image and text content from multimodal inputs
    • Weights under the MIT license
    Source: GLM-4.6V: Open Source Multimodal Models with Native Tool Use
  40. MeituanImageOpen weights

    LongCat-Image

    LongCat-Image is Meituan's open 6B-parameter image generation model, released together with a developer checkpoint and the LongCat-Image-Edit image editing model.

    Modalities
    text, image → image
    Parameters
    6B
    License
    Apache-2.0
    At launch
    Open weights
    • 6B parameters
    • LongCat-Image, LongCat-Image-Dev, and LongCat-Image-Edit weights released together
    • A distilled LongCat-Image-Edit-Turbo followed on February 3, 2026
    • Licensed under Apache-2.0
    Source: GitHub - meituan-longcat/LongCat-Image
  41. GoogleReasoningProprietary

    Gemini 3 Deep Think

    Gemini 3 Deep Think is an enhanced reasoning mode built on Gemini 3 Pro for complex math, science and logic problems, rolled out to Google AI Ultra subscribers in the Gemini app.

    Modalities
    text → text
    At launch
    Consumer app
    • Introduced alongside Gemini 3 on Nov 18, 2025 as an enhanced reasoning mode coming soon to Ultra subscribers
    • Received a major upgrade for science, research and engineering on Feb 12, 2026
    • The February 2026 upgrade opened Gemini API access to select researchers and enterprises through an early access program
    Source: Gemini 3 Deep Think is now available
  42. Nous ResearchReasoningOpen weights

    Hermes 4.3

    Update to the Hermes series that nearly matches Hermes 4 70B at half the size, post-trained entirely on Nous Research's decentralized Psyche network.

    Modalities
    text → text
    Context
    500K context
    Parameters
    36B
    License
    Apache-2.0
    At launch
    Open weights
    • Trained with extended context up to 512K tokens
    • Released as Hermes-4.3-Seed-36B
    • Psyche-trained version outperformed a centrally trained version on downstream tasks
    Source: Introducing Hermes 4.3: Local Intelligence Globally Trained | Nous Research
  43. KuaishouVideoProprietary

    Kling Video 2.6

    Kling Video 2.6 is Kling's first model with simultaneous audio-visual generation, producing visuals together with voiceovers, sound effects, and ambient sound in a single pass from text or image prompts.

    Modalities
    text, image → video, audio
    At launch
    Consumer app
    • Generates speech, sound effects, and ambient audio with the video
    • Supports Chinese and English voice generation
    • Generates clips up to 10 seconds long
    Source: Kling AI Launches Video 2.6 Model with "Simultaneous Audio-Visual Generation" Capability, Redefining AI Video Creation Workflow
  44. ByteDance SeedImageProprietary

    Seedream 4.5

    Seedream 4.5 is an upgrade to Seedream 4.0 that scales up the model to improve subject consistency in multi-image editing, preserve reference image details, and strengthen typography and dense text rendering.

    Modalities
    text, image → image
    At launch
    API
    • Supports up to 10 reference images for multi-image fusion
    • Stylized output up to 4K resolution
    • Launched in open beta on BytePlus ModelArk
    Source: BytePlus Unveils Seedream 4.5: Precision-Focused Upgrade Delivering Sharper Visuals, Smarter Control, and 4K Creative Consistency
  45. AmazonReasoningProprietary

    Amazon Nova 2 Lite

    Amazon Nova 2 Lite is a fast, cost-effective reasoning model for everyday workloads, with optional extended thinking at three intensity levels, multimodal input and a one-million-token context window.

    Modalities
    text, image, video → text
    Context
    1M context
    At launch
    API
    • Extended thinking is off by default, with low, medium and high levels
    • One-million-token context window and up to 64K output tokens
    • Built-in web grounding and code interpreter tools
    Source: Introducing Amazon Nova 2 Lite, a fast, cost-effective reasoning model
  46. AmazonSpeechProprietary

    Amazon Nova 2 Sonic

    Amazon Nova 2 Sonic is Amazon's second-generation speech-to-speech model for conversational AI, adding Portuguese and Hindi, polyglot voices, configurable turn-taking, crossmodal text and voice input, and asynchronous tool calling.

    Modalities
    text, audio → text, audio
    Context
    1M context
    At launch
    API
    • Supports seven languages: English, French, Italian, German, Spanish, Portuguese and Hindi
    • One-million-token context window per the Bedrock model card
    • Integrates with Amazon Connect and telephony providers
    Source: Introducing Amazon Nova 2 Sonic: Our new speech-to-speech model for conversational AI
  47. PerplexityLLMOpen weights

    BrowseSafe

    An open prompt-injection detection model fine-tuned from Qwen3-30B-A3B that reads a web page's HTML and answers whether it contains malicious instructions aimed at a browser agent, released with the BrowseSafe-Bench benchmark.

    Modalities
    text → text
    Context
    16K context
    License
    MIT
    At launch
    Open weights
    • Outputs a single yes or no token for each page
    • Context length up to 16,384 tokens per the model card
    • BrowseSafe-Bench spans 11 attack types and 9 injection strategies
    • Built for Perplexity's Comet browser; weights on Hugging Face under the MIT license
    Source: Building Safer AI Browsers with BrowseSafe
  48. Mistral AISmall modelOpen weights

    Ministral 3 (3B, 8B, 14B)

    Family of three dense small models released with Mistral Large 3, each with image understanding and offered in base, instruct and reasoning variants.

    Modalities
    text, image → text
    Context
    250K context
    Parameters
    3B, 8B and 14B
    License
    Apache-2.0
    At launch
    Open weights, API
    • 256K-token context window
    • Optimized for edge hardware including DGX Spark, RTX PCs and Jetson
    • API model IDs ministral-3b-2512, ministral-8b-2512 and ministral-14b-2512
    Source: Introducing Mistral 3 | Mistral AI
  49. Mistral AIMultimodalOpen weights

    Mistral Large 3

    Mistral's first mixture-of-experts model since Mixtral, a sparse open-weight multimodal model trained from scratch on 3,000 NVIDIA H200 GPUs and released under Apache 2.0.

    Modalities
    text, image → text
    Context
    250K context
    Parameters
    MoE: 675B total, 41B active
    License
    Apache-2.0
    At launch
    Open weights, API
    • Base and instruct versions released
    • NVFP4 checkpoint for Blackwell NVL72 and single 8-GPU nodes
    • API model ID mistral-large-2512
    Source: Introducing Mistral 3 | Mistral AI
  50. DeepSeekReasoningOpen weights

    DeepSeek-V3.2

    DeepSeek-V3.2 is the official successor to V3.2-Exp, a reasoning-first model built for agents. It is DeepSeek's first model to integrate thinking directly into tool use, supporting tool calls in both thinking and non-thinking modes.

    Modalities
    text → text
    License
    MIT
    At launch
    API, Consumer app, Open weights
    • Built on DeepSeek Sparse Attention for efficient long-context inference
    • Trained with agent data synthesized across more than 1,800 environments and 85,000 complex instructions
    • Powers both deepseek-chat (non-thinking) and deepseek-reasoner (thinking) on the API
    Source: DeepSeek-V3.2 Release
  51. DeepSeekReasoningOpen weights

    DeepSeek-V3.2-Speciale

    DeepSeek-V3.2-Speciale is a high-compute reasoning variant of DeepSeek-V3.2 that pushes maximum reasoning capability at the cost of higher token usage. DeepSeek reports gold-level results at IMO, CMO, the ICPC World Finals, and IOI 2025.

    Modalities
    text → text
    License
    MIT
    At launch
    API, Open weights
    • Served through a temporary API endpoint available until December 15, 2025
    • No tool calling support at launch, offered for community evaluation and research
    • Weights on Hugging Face under the MIT License
    Source: DeepSeek-V3.2 Release
  52. RunwayVideoProprietary

    Gen-4.5

    Video generation model advancing pre-training data efficiency and post-training, with improved motion, physical accuracy, prompt adherence and temporal consistency at Gen-4's speed and efficiency.

    Modalities
    text, image → video
    At launch
    Consumer app
    • Available for all paid plans from December 11, 2025
    • Image to video added January 21, 2026
    • Trained and served on NVIDIA Hopper and Blackwell GPUs
    Source: Runway Gen-4.5: State-of-the-Art AI Video Generation
  53. KuaishouVideoProprietary

    Kling O1

    Kling O1 is a unified multimodal video model that combines reference-based generation, text-to-video, start and end frames, in-painting, restyling, and shot extension in one engine built on a Multimodal Visual Language framework.

    Modalities
    text, image, video → video, image
    At launch
    Consumer app
    • Accepts text, image, video, and subject references as prompts
    • Supports conversational video editing without manual masking
    • Kling Image O1 image model followed on December 2, 2025
    Source: Kling O1 Launches as the World's First Unified Multimodal Video Model

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 December 2025." The AI Engineering, updated Dec 31, 2025. https://theaiengineering.com/releases/2025/12

HTML link

<a href="https://theaiengineering.com/releases/2025/12">AI models released in December 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