wrapped·Jul 01, 2026
/r/r_u4iojm6-ekgyour run
Yi-Coder-9B-Chat
on RTX 5090
199.3
tok/sdecode
RTX 5090 (32GB)
rank in tier
42/208RTX 5090 runstop 20%
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
- gemma3 on RTX 4090195.0 tok/s
- stable-code-instruct-3b-4bit on M3 Ultra192.5 tok/s
- coder-v2 on RTX 3090189.5 tok/s