vLLM's study also concluded that "FP8 can deliver meaningful latency and capacity gains with small or negligible accuracy loss". Their benchmarks include LiveCodeBench 6.

https://vllm-project.github.io/2026/04/22/fp8-kvcache.html