According to CoinMeta, MiniMax has today open-sourced the weights of the M3.1 Flash model, which have been made available on Hugging Face and Magada platforms. This model adopts a MOE architecture with a total of 428 billion parameters, but only about 23 billion parameters are activated during each generation. The new version fully utilizes sparse attention mechanisms and incorporates quantized training as well as Dspark inference decoding techniques. Officials claim that the cost of performing Agent tasks has been reduced by 70% compared to the previous generation, and the generation speed has nearly tripled. In terms of performance, it scored 73.7% on the Deepswe V1 .1 benchmark, which is lower than the 77% achieved by the previous generation's Deepseek V4 .1 and Flash-1017. An independent evaluation agency, Vals AI, found that its programming capabilities have significantly improved compared to the previous generation, but it still struggles with more complex tasks. The model uses MiniMax Community License for its weights. For commercial use, it is necessary to indicate the source of the model; products or services that generate revenue of over 20 million US dollars annually require prior written authorization.