The Fastest Kimi K3 Kernels in vLLM Were Written by Agents

The interesting number in vLLM’s Kimi K3 work is not the speedup. It is who wrote the code.

• Coding agents built a unified CuTe DSL decode path for Kimi K3, covering attention, MoE and communication across 8 or 16 Blackwell GPUs.
• One transformer layer went from 23 kernels and 130 microseconds to 13 kernels and 69 microseconds.
• Gains run from 1.88 times at batch one on 16 GB300 GPUs down to 1.62 times at batch 64, so the whole curve moves rather than a single point.
• With DSpark the same 16 GPU setup clears 842 tokens per second at batch one, 1.73 times stock vLLM under identical speculation.
• The team proposed a fast-track RFC that keeps each optimization inside the model’s code and requires published correctness and performance evidence before a maintainer reviews it.

The review process, not the kernel, is the constraint now, and that is the harder problem to buy your way out of. Read the full analysis.