Kimi K3 is the largest model we feature, with 2.8T total and 104B active parameters. MI355X serves ~8M tokens/s per MW at 50 tok/s per user, about a third of B300 (~24M), which uses NVIDIA's multi-node disaggregated configs. Its curve is shallow, peaking at about 13M at 12 tok/s per user, and for a single agent it reaches 152 tok/s against 235 on B300.
