Running Qwen3.8 on a Steam Machine with 8GB VRAM: ~18 t/s decode at 128K context

Accurate-Door3692 · reddit · 2026-11-07

A Redditor benchmarked Qwen3.8-Flash-Next Gyro S on a stock-SteamOS Valve Steam Machine (AMD Navi 33, 8GB VRAM + 32GB RAM), hitting 235 t/s prefill and 15-18 t/s decode at 128K context.

Key findings:

--set resident-budget-gib=64 was the only tweak that helped (+20% decode) by pinning experts in the engine's own RAM tier instead of the OS page cache.
pcie-frac, larger kv-resident, and bigger VRAM reserves showed no measurable gain; the quick tuner did nothing.
Manually enabled MTP speculative decoding gave no speedup and cut prefill 10% — not worth it on 8GB VRAM. Also, --mtp on silently fails: the draft layer is never downloaded.
SteamOS tips: SSH-launched processes die with the session (use systemd-run), and always check the generated config to see which flags actually applied.

The author warns only 3.5-4GB RAM stays free and the machine rebooted once during testing; all scripts are open-sourced on GitHub.

Original post →

Running Qwen3.8 on a Steam Machine with 8GB VRAM: ~18 t/s decode at 128K context

More from Infra