wrapped·Jul 01, 2026
/r/r_a56-wxl21lkyour run
Qwen3.6-35B-A3B-Q4_K_M.gguf
on RTX 5090
224.0
tok/sdecode
RTX 5090 (32GB)
rank in tier
17/208RTX 5090 runstop 8%
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
- Sage-A1.5-ctx48k on RTX 4090217.2 tok/s
- gpt-oss-20b on RTX 3090207.2 tok/s
- Qwen3-Coder-30B-A3B-Instruct on RTX 3090199.7 tok/s