qwen2.5-coder

on RTX 3090

69.2

tok/sdecode

RTX 3090 (24GB) + AMD EPYC 7702P 64-Core Processor (64c) + 252GB

rank in tier

4/4RTX 3090 runs

best workload

chat-short

where the rig flew

slowest workload

agent-trace63.6 tok/s

where the rig struggled

backend

ollamaollama is llama.cpp under the hood; pure llama.cpp tends to nudge a few % faster.

faster than

starcoder2-15b-4bit on M3 Ultra64.2 tok/s
Codestral-22B-v0.1-4bit on M3 Ultra47.5 tok/s
Qwen2.5-32B-Instruct-4bit on M3 Ultra34.6 tok/s

This is a shareable view of one signed llm-speed suite-suite-v1 run. Numbers are not edited. The raw record has every workload, the public key, and the fingerprint hash.