MiniMax
MiniMax releases M3.1-Flash weights under a community license
MiniMax posted M3.1-Flash weights with a license that requires credit for commercial use and written permission above $20 million in yearly revenue.
MiniMax released the weights for MiniMax-M3.1-Flash, a model with about 428 billion parameters, and attached a license that makes commercial users credit the model and the largest ones ask permission first.
MiniMax posted the weights on Hugging Face, and the model card is dated Oct. 23, 2026. The model is generally available, with the API live and the weights downloadable. OpenRouter first listed it Oct. 19.
What Changed From M3
M3.1-Flash builds on MiniMax-M3 and changes how the model runs, according to the card. It is a native multimodal mixture-of-experts model for coding, agentic workflows, computer use and cybersecurity. It takes text, image and video as input and outputs text. It has about 428 billion total parameters, 23 billion active per token and a context window of up to 1 million tokens.
The first three full-attention layers are now sparse attention. The routed experts and attention paths were trained with quantization in mind. The card says this reduces memory use and improves serving efficiency. DSpark speculative decoding replaces the earlier EAGLE-like approach, and the release includes the draft head for compatible serving software.
MiniMax recommends SGLang, vLLM and TokenSpeed for serving, with temperature 1.0 and top_p 0.95.
Scores
The benchmark figures are MiniMax’s own, from the model card dated Oct. 23, 2026, and are not independent results. The card says they apply to the official weights and that third-party quantizations may differ.
On BankerToolBench, M3.1-Flash scored 76.2, tying Opus 5 at 76.2. GPT-5.6 Sol scored 74.6, GLM-5.3 74.5, DeepSeek V4.1 Flash 73.7 and Kimi K3 69.1. On Automation-Bench v1.0.6 it scored 62.8, against 60.5 for DeepSeek V4.1 Flash-1017, 50.3 for Opus 5 and 45.8 for GPT-5.6 Sol.
It trails on two evals. Terminal Bench 4.0 gave it 38.1, behind Opus 5 at 51.8, DeepSeek V4.1 Flash-1017 at 42.3 and GPT-5.6 Sol at 39.9, and just ahead of GLM-5.3 at 37.9. On GDP.pdf AA it scored 20.8, against 27.2 for GPT-5.6 Sol and 21.6 for Opus 5.
The card also lists 73.7% on DeepSWE v1.1 and 82.2 on CyberGym.
License
The MiniMax Community License is not an OSI open-source license. It is free for non-commercial use. Commercial use requires displaying “Built with MiniMax M3.1-Flash” on a related website, interface, blog post, about page or documentation.
A product or service that generates more than $20 million in yearly revenue needs prior written authorization from MiniMax. Others send a one-time notice by email. Commercial use includes paid services built on the model, commercial API use and commercial deployment of fine-tuned or modified versions. An appendix bars uses including military purposes. The terms come from the LICENSE file in the same repository, read Oct. 23, 2026.
Pricing
MiniMax’s pay-as-you-go pricing page, checked Oct. 23, 2026, lists M3.1-Flash at $0.15 per million input tokens, $0.60 per million output tokens and $0.015 per million for cached prompt reads. OpenRouter lists the same $0.15 and $0.60 rates, with a 1,048,576-token context window.
Our View
The $20 million line decides which companies building paid products can use these weights without asking. A company under it sends MiniMax an email.
A company over it needs MiniMax’s written authorization first. It cannot put the model in a paid product before then.
The benefit holds for as long as a product stays under $20 million in yearly revenue, or until MiniMax grants the authorization.