MiniMax has open-sourced MSA—million-token context inference no longer requires burning tokens. The MIT license allows direct commercial use. This week, I’m waiting for the M3 weights.

CryptoWorld reports that MiniMax has announced its high-performance attention library MSA for NVIDIA Blackwell (SM100) is now open source, licensed under MIT, with MiniMax-M3 weights released this week. MSA filters KV blocks, performing attention only on relevant blocks, supporting million-scale context inference. Research shows that in a 1 million token scenario, MSA significantly reduces computation compared to Dense GQA and achieves notable acceleration on H800. The open-source version integrates C++ JIT and Cute-DSL, supporting multiple precisions.