On June 1, Xiyu Technology officially released the new generation model MiniMax M3. This model has cutting-edge programming capabilities, up to 1M ultra-long context, and supports native multi-modality (pictures, video input, and computer desktop operations). It has become the first model in China to have these three capabilities at the same time, and is currently the only open source model.
According to official disclosures, on the programming evaluation set SWE-Bench Pro, M3 scored 59.0%, surpassing GPT-5.5 and Gemini 3.1 Pro, and close to Opus 4.7; on the Agent evaluation Claw-Eval, M3 received the highest score; on the multi-modal test set OmniDocBench, M3 scored higher than Gemini 3.1 Pro.
M3 adopts a new sparse attention architecture MSA (MiniMax Sparse Attention). Under 1 million contexts, the calculation amount per token is only 1/20 of the previous generation model. The prefilling stage is accelerated by more than 9 times, and the decoding stage is accelerated by more than 15 times.
MiniMax has simultaneously updated the Agent product MiniMax Code and launched a Token Plan subscription plan (Plus 49 yuan/month, Max 119 yuan/month, Ultra 469 yuan/month). The M3 API is open for use from now on, and the 512k context version is available at a 50% discount for a limited time, lasting for 7 days. Model weights and technical reports will be open source within 10 days.