MiniMax M3.1 Flash weights: a 1M-context agent and coding model with 23B active parameters
What is it?
MiniMax M3.1 Flash is a multimodal MoE model aimed at coding, agentic work, computer use, and security tasks. It has about 428B total parameters with 23B active per token, accepts text, image, and video, and handles up to 1M tokens of context. It uses sparse attention throughout, was trained with quantization in mind, and ships with a DSpark draft head for speculative decoding. You can serve it with SGLang or vLLM.
hf download MiniMaxAI/MiniMax-M3.1-Flash --local-dir MiniMax-M3.1-FlashWhy does it matter?
MiniMax cites 61.7 on OSWorld 2.0Vendor claim, 73.7% on DeepSWE v1.1Vendor claim, and 76.2 on BankerToolBenchVendor claim. Its own table also shows 38.1 on Terminal Bench 4.0Vendor claim, below DeepSeek V4.1 Flash-1017 (42.3)Vendor claim. With only 23B active parameters it should be cheaper to serve than models in the same performance band, but the license is the MiniMax Community License, not open source, so read the terms before any commercial use.
Maturity
GAThe weights are ungated and downloadable now. The license isn't open source, and Transformers support is still in progress.
How to track it
- Independent numbers: look for M3.1 Flash on Artificial Analysis and the official OSWorld and SWE-style leaderboards.
- Local use: check whether the Transformers, unsloth and KTransformers support listed as in progress lands.
- License: watch the repo's LICENSE file for changes to the MiniMax Community License's commercial terms.