Report
Kimi K3.1 Pro reportedly hits 300 tokens per second on eight AMD MI350X GPUs
The reported setup used the default SGLang recipe, with roughly 60 tokens per second as the K3 comparison.
TLDR
Kimi K3.1 Pro reportedly ran at 300 tokens per second on eight AMD MI350X GPUs using the default SGLang recipe, versus roughly 60 for K3. The account behind the report preferred it over Astra or Opus in some cases because of its speed. This was a firsthand account of performance and a speed-based preference, not a model-quality ranking.
Combined views
—
3 Sources, first seen ago
