MiniMax-M3.1 · decode, M ≤ 32

One MiniMax-M3.1 sparse layer, kernel by kernel

Blackwell (SM100), TP decode at ≤32 tokens: MXFP8 linears, NVFP4 W4A4 routed experts, NVFP4 KV and index-K caches. One block = one kernel launch; stacked blocks stand for several. Numbers mark this post's optimizations; click one to highlight its blocks.