MiniMax M3 发布：长上下文多模态模型，支持文本/图像/视频推理
多模态推理模型终于有了高效的开源选择——MiniMax M3 用 23B 激活参数实现长上下文多模态推理，做视频分析或长文档处理的团队可以直接在 NVIDIA 端点免费试，值得关注。
MiniMax 团队发布了 MiniMax M3，这是一个支持文本、图像和视频推理的长上下文多模态模型。模型采用稀疏注意力机制，总参数量约 428B，激活参数仅约 23B，在保持高性能的同时大幅降低了计算成本。
该模型已开源权重，可在 Hugging Face 获取，并可通过 NVIDIA 的 GPU 加速端点免费试用。M3 的长上下文能力使其在处理视频、长文档等场景中具有优势。
Congrats to the @MiniMax_AI team on the release of MiniMax M3, a long-context multimodal model for text, image, and video reasoning. 🙌 Try it today with our free GPU-accelerated endpoint on build.nvidia.com 0. Details: nvda.ws/4v4BWhD W MiniMax (official) @MiniMax_AI MiniMax M3, Open-Weight, Now On Hugging Face , with only ~428B parameters and ~23B activated parameters Weights: huggingface.co/MiniMaxAI/Mini… MiniMax Sparse Attention: huggingface.co/papers/2606.13… 🔗 View Quoted Tweet
This report is categorized under AITOP's "Models" desk, focusing on China's AI industry and domestic foundation model ecosystem. Within our cross-regional analytical framework, it directly maps to US-China foundation model benchmarks and commercialization comparison. Key entities and technical topics under continuous tracking include MiniMax, M3, 多模态模型, 长上下文.
Source attribution: curated from NVIDIA AI (@NVIDIAAI). Cross-source corroboration: 11 independent sources have covered this development (including vLLM, LMSYS Org (SGLang), Simon Willison’s Weblog), confirming industry-wide signal strength.
- vLLM06-12 14:47Source
- LMSYS Org (SGLang)06-12 14:18Source
- Simon Willison’s Weblog06-10 20:00Source
- Richard Socher06-11 15:30Source
- Sebastian Raschka06-12 04:42Source
- Decoder06-10 19:20Source
- Together AI06-11 20:04Source
- lmarena.ai06-12 20:28Source
- ollama06-13 01:26Source
- Tri Dao (FlashAttention)06-12 04:20Source
