MiniMax M3 gets day-zero vLLM support, 1M context window and MSA sparse attention
Original title:day-0 in @vllm_project and it comes with: dedicated MSA prefill/decode kernels, 1M-context serving ...
AI summary
MiniMax M3 is released, with frontier coding and agent capabilities, native image and video input, computer use, and a 1M-token context. At its core, it uses MSA sparse attention: each query scores 128-token KV blocks and attends only to top blocks. vLLM supported M3 on the same day, including dedicated MSA prefill/decode kernels, prefix caching and chunked prefill, BF16 and MXFP8 checkpoints, MoE backends for Hopper and Blackwell, and validation on NVIDIA and AMD hardware. It also supports native multimodal input, tool calling, reasoning parsing, and thinking mode control for agent workloads.
Key Takeaways
- MiniMax M3 is released,with frontier coding and agent capabilities,native image and video…
- At its core,it uses MSA sparse attention: each query scores 128-token KV blocks and attends only…
- vLLM supported M3 on the same day,including dedicated MSA prefill/decode kernels,prefix caching…
- It also supports native multimodal input,tool calling,reasoning parsing,and thinking mode control…
Quick Facts
- Published
- Updated
- Source
- X:MiniMax (@MiniMax_AI)
- Category
- Model Releases
- AI score
- 82
M3把1M上下文从‘理论上能做’变成了‘今天就能部署’,MSA稀疏注意力是关键,开源社区和推理框架的深度合作值得关注。
Related entities
Source & copyright
This article is based on curated titles and AI summaries from the AIHOT public API, with attribution and links back. Original copyright remains with the original publishers and authors; this site does not mirror full texts.
How to Cite
When citing this article, keep the title, date, and permalink, and credit Orange News as the source.
Orange News: MiniMax M3 gets day-zero vLLM support, 1M context window and MSA sparse attention. June 13, 2026. https://new.nat6.net/en/news/cmqbgej4i019vslamev56gffo/