Physix Frontier · News Briefing Card (IT Home · Oct 18, 2026)

Nvidia Optimizes Kimi K3 Inference, 1.68x Faster on Same Hardware

KEY FACTS

SemiAnalysis reports that Nvidia made upstream optimizations to vLLM, boosting Kimi K3 inference speed.
Kimi K3 has 2.8 trillion parameters and reaches 423 tok/s/user after optimization.
On the same GB300 NVL72 hardware, performance improved 1.68x over the previous version.
The optimization was done by Benjamin, an engineer at Nvidia CentML.
The optimization uses Megakernel fusion, with roughly 12,000 lines of kernel code.

KEY DATA

2.8 trillionKimi K3 parameters

423 tok/s/userInference speed

1.68xPerformance gain

12,000 linesMegakernel code size

PHYSIX OBSERVATION

This speedup is not about swapping chips; it squeezes more throughput out of existing GB300s through kernel fusion, showing that large-model inference competition is shifting from hardware stacking to software engineering. For users, the same compute can serve more people, and inference costs are likely to fall; for the industry, upstream vLLM optimization is becoming Nvidia's new lever for locking in its ecosystem.

Source: IT Home report