Pull requests / #1998

#1998 bench: community suite 1 (--long) on RX 9070 XT, engine 0.1.43, IQ3_S

open · @Thxeverybody · 0 コメント · GitHub で見る

BenchmarksSetup & installModels & quantsLinux

本文

## Title
bench: community suite 1 (--long) on RX 9070 XT, IQ3_S

## Summary
Community benchmark, suite 1 with `--long`, run on 2026-10-11 — first IQ3_S entry on RDNA4 that I can find: RX 9070 XT (gfx1201, 16 GiB) + Ryzen 7 9700X + 128 GB DDR5, Linux (CachyOS), engine 0.1.43, config as measured by `setup.sh --calibrate` plus `--vram-reserve-mib 500` (desktop on iGPU).

32K fresh prompt: 2726 tok/s median (range 2723.5-2726.9), first token after 11.8 s. Agent session: 68.8% cache hits, ~3 re-read tokens per session. Chat decode 65-68 tok/s.

Happy to re-run with different settings if a specific comparison would be useful.

関連リンク

インストール・モデル・リリースへの站内リンク。