MiniMax M3.1 Flash Open-Source: 428 Billion Parameters, Agent Task Costs Cut by 70%

Beating AI News Flash: MiniMax today open-sourced the weights of its M3.1 Flash model, which is now available on Hugging Face and ModelScope. Space Bunny, which launched anonymously earlier and reportedly processed over 200 trillion tokens total, is an early iteration of M3.1 Flash. The model uses a MoE architecture with 428 billion total parameters, but only activates around 23 billion per generation. The new version fully adopts sparse attention, and adds quantization training and DSpark speculative decoding. The company states that Agent task costs are 70% lower than the previous generation, with generation speed nearly tripled. In terms of performance, it scores 73.7% on DeepSWE v1.1, lower than DeepSeek V4.1 Flash-1017's 77%. Independent evaluator Vals AI noted its programming capabilities have improved significantly over the prior version, though complex terminal tasks remain a weakness. The weights are released under the MiniMax Community License. Commercial use requires attribution of the model source; if the annual revenue of related products or services exceeds $20 million, prior written authorization is required.