Kimi-K3’s MXFP4 weights leave only ~50 GiB free on an 8x MI300X node – barely enough to serve a 1.4 TiB model. This entry covers the disaggregated, wide expert-parallel setup (2P2D, EP16 in vLLM) that makes it work, plus how MXFP4 runs on CDNA 3 without a native MXFP4 instruction, and two concurrency bugs we found and fixed before trusting any benchmark number.
Useful if you’re taking a frontier MoE model to multi-node production – you get a working layout and a capacity model that turns an SLA into a node count before you buy hardware.
