Qwen Model Runs on Valve Steam Machine at About 18 Tokens per Second
A Reddit user running Qwen3.8-Flash-Next Gyro S on a Valve Steam Machine with an 8 GB AMD Navi 33 GPU reported about 18 tokens per second decoding and 235 tokens per second prefilling in Strata. Setting the resident budget to 64 GB improved decoding by roughly 20%, while a manually loaded MTP draft layer produced no gain and reduced prefilling by about 10%. The figures came from three runs per configuration on one system and were not independently verified.
The report provides a detailed local-inference configuration and tuning baseline for running a large model on an 8 GB consumer GPU, although the results apply only to the tester's hardware and setup.
I followed up on the post about running Qwen 3.8 Flash Next with 8 GB VRAM and 32 GB RAM. I wanted to see how it does on my Steam Machine.
Setup
- Valve Steam Machine, stock SteamOS 3.9.5, AMD Navi 33 with 8 GB, 32 GB RAM, NVMe
- Stock Strata terminal 0.2.9 (engine 0.1.55, HIP package), Gyro S TQ1_0
- 128K context, q4_0 KV, no vision, MTP off, served on the LAN
Results (Strata's own timing line from the engine log, about 8K-token prompts, 3 runs each)
| Config | Prefill t/s | Decode t/s |
|---|---|---|
Default (--tune skip) | 219-240 | 15.1-15.9 |
| Quick tuner | 237-239 | 14.5-17.2 (tuner said "already fastest") |
pcie-frac 0.35 | 239 | 14.8-16.0 |
kv-resident 20480 + vram-reserve-mib 900 | 227-229 | 14.8-15.5 |
resident-budget-gib 64 | 232-242 | 17.3-19.8 |
#Valve#Steam Machine#SteamOS#Qwen3.8-Flash-Next Gyro S#Strata#AMD Navi 33