wrapped·Nov 12, 2026
/r/r_4nge7d__8-xyour run
Qwen3.8-27B
on RTX 5090
30.3
tok/sdecode
RTX 5090 (32GB)
rank in tier
203/208RTX 5090 runstop 98%
best workload
concurrent-decode
where the rig flew
slowest workload
—
single-workload run
backend
llama.cppllama.cpp is the universal backend — broad model support, modest speed ceiling.
faster than
- Llama-3.1-8B-Instruct-4bit on M3 Pro29.2 tok/s
- Qwen3.8-27B-4bit on M3 Ultra27.3 tok/s
- Muse-Spark-Open-816B-A42B-UD-00001-of-00005 on 3x A10026.3 tok/s