MiniMax open-sources M3, a million-token model it pegs level with GPT-5.5 on SWE-Bench Pro

Chinese lab MiniMax has published the open weights for M3, an agentic coding model that scores 59.0% on SWE-Bench Pro, which MiniMax puts on par with GPT-5.5, while holding a one-million-token context. The weights are now on Hugging Face after a phased rollout from the 1 June launch.

The unlock is attention, not raw scale. M3 is a sparse Mixture-of-Experts model, and its "MiniMax Sparse Attention" is what makes a million-token window cheap enough to actually serve rather than a benchmark stunt. It is also natively multimodal, reading images and video in the same checkpoint, and reports 66.0% on Terminal Bench 2.1.

An openly downloadable checkpoint that holds a million tokens and codes near the closed frontier chips away at the per-seat pricing of proprietary tools. The catch is the usual one: a SWE-Bench Pro number set in a harness is not the same as performance on a real, messy repository, and open weights still need someone with the GPUs to run them.