[MiniMax Open-Sources Blackwell-Specific Attention Library, M3 Weights Expected to Be Released This Friday] According to monitoring by Beating, MiniMax Developer Relations Lead Ryan Lee announced that the high-performance attention library MiniMax Sparse Attention (MSA), designed for NVIDIA Blackwell (SM100) GPUs, has been officially open-sourced under the MIT license. Ryan Lee also stated that the MiniMax-M3 weights are expected to be released this Friday. MSA has been applied to MiniMax-M3's million-token context inference, where it selects the most relevant KV Blocks within each GQA group and performs attention computation only on the selected blocks. Research shows that under a million-token context, MSA reduces attention computation by 28.4 times compared to Dense GQA with the same configuration, achieving 14.2x prefill acceleration and 7.6x decoding acceleration on H800 GPUs. The open-source version integrates two implementations—C++ JIT and CuTe-DSL—within the same Python package, and also provides Dense FlashAttention and Sparse Top-k Attention Kernels, supporting multiple precision formats such as BF16, FP8, NVFP4, and FP4. It is currently primarily targeted for deployment on NVIDIA Blackwell (SM100) GPUs. [Original Link]