Evidence record / source view

RTX 3090 · Qwen3.8 27B

Published conditions are locked before values enter the comparison table.

ProfileOrdinary local median

Runtimellama.cpp · disclosed build

QuantQ4 class · source disclosure

Concurrencysingle user

ScopeQwen3.8 27B

Median reading705prefill · tok/s · source median

Median writing65decode · single stream · tok/s · source median

Ordinary source points1reading n=1 · writing n=1

Last reviewed23 Aug2026

Performance plane

Reading and writing stay separate.

tok/s · higher is better

included ordinary source excluded experimental path per-metric median

A point appears only when both reading and writing values are disclosed for the same run.

Source-observation spread

Keep the source points visible.

included ordinary source excluded experimental path medianscale unavailable

Compatible values are never silently merged with a different configuration.

Evidence ledger

Conditions stay attached to the value.

2 disclosed observations

Result	Conditions	Disposition	Source / date

Method note

Benchmark values describe source observations, not a guarantee of your machine’s performance. Reading (prefill) and writing (single-stream decode) are medianed independently across normal, deployable Qwen3.8 27B runs. Stable MTP is ordinary; SGLang and DFlash/DFlash2 paths are retained outside this default view until they have a comparable deployment rule.