vLLM blog: agent-assisted optimization of Kimi K3 decoding on NVIDIA GB300

The vLLM project blog, in a post by an NVIDIA engineer dated Oct 16, 2026, describes a unified CuTe DSL decode path for Kimi K3 on NVIDIA GB300 GPUs. The authors used coding agents to write much of it.

• The post claims over 840 tokens/s at batch 1 on 16 GB300 GPUs with DSpark speculation (synthetic acceptance length 6). It also claims 1.5–2x higher decode throughput in several tested configurations and up to 1.8x higher P90 interactivity on the AgentX workload. These are the authors' benchmarks.
• It also proposes a fast-track RFC for model-specific optimizations, aimed at reducing review burden because agent-written kernels are outpacing maintainers' review capacity. Kimi K3 is the first contribution made under that approach.