Blog

Measured, not marketed.

Field notes on LLM inference performance, from Verinfer's own runs.

October 30, 2026 · Quantization

FP8 Cut Our TPOT by 25%. At High Concurrency, It Also Changed the Answers.

Decoding got 23–26% faster per token. At 64 concurrent requests, self-consistency fell from 97.6% to 75%, and here is why.