MiniMax puts M3.1 Flash weights online with a $20 million limit
A company with more than $20 million in annual revenue must get authorization from MiniMax to use the weights.
On 23 October 2026, MiniMax put the weights of its M3.1 Flash model on Hugging Face, four days after launch. The files are 252.4 GB, and 23 billion of the 428 billion parameters work on each token. The weights are open, but a company with more than $20 million in annual revenue must get authorization from MiniMax.
How it works
The model is a mixture of experts. This shows that only some parts of the model work on each token. It reads text, image and video, and it writes text.
The numbers
The vLLM tool reports 76 tasks at one time on four NVIDIA B300 GPUs. Each task has 256K tokens. The older M3 model fits 28. These are tool tests on set hardware, not tests in use.
What is not known
The license is not open source. It also makes a company show an attribution notice on each business use. The API costs $0.60 for each million output tokens. Four B300 GPUs cost more to rent if a team does not keep them busy. Other groups did not check the tests.
Insight: Free trial first, license limit second
On 19 October, a gateway offered M3.1 Flash free for a short time, and MiniMax gave no test numbers. Four days after, the weights came with a $20 million revenue limit. Mistral will release Large 4 open weights on 31 October and gave no license terms. This reads as a race where open weights get users first and the license sets the price after. Monitor the Mistral terms.
Sources
Posted
