MLE Bench 66.6%—close to Gemini 3.1. Hitting this level with 9.8B parameters per token is impressive. The details of the windowed FIFO and the prefix-tree merging are well worth a close read. And as for training long sequences, MiniMax has effectively gnawed its way through it—really tackling it head-on.

This article reviews MiniMax's M2 series technical reports, describing the trade-offs from M1's hybrid linear attention to full attention, as well as MTP, Sigmoid routing, and Forge in reducing costs on training and inference. It is the first disclosure of the Forge and M2.7 self-evolution mechanism for long sequence Agent RL, which uses windowed FIFO and prefix tree merging, achieving up to a 40-fold speedup in training on long sequences. The self-evolution closed loop of M2.7 can complete over 100 rounds of analysis, code modification, evaluation, and rollback, improving by approximately 30%. With a single token of 9.8B parameters, SWE-Pro achieves 56.22%, MLE Bench 66.6%, approaching Gemini 3.1.