wrapped·Nov 16, 2026
/r/r_exg-o5xt410your run
Qwen3.8-27B
on RTX 5090
71.3
tok/sdecode
RTX 5090 (32GB)
rank in tier
104/208RTX 5090 runstop 50%
best workload
chat-short
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
- Qwen2.5-14B-Instruct-4bit on M3 Ultra70.9 tok/s
- Qwen2.5-Coder-14B-Instruct-4bit on M3 Ultra70.5 tok/s
- starcoder2-15b-4bit on M3 Ultra64.2 tok/s