r1

on RTX 4090

133.8

tok/sdecode

RTX 4090 (24GB) + AMD EPYC 75F3 32-Core Processor (64c) + 504GB

rank in tier

13/31RTX 4090 runstop 42%

best workload

agent-trace

where the rig flew

slowest workload

chat-long131.9 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

Qwen2.5-14B-Instruct on RTX 5090133.3 tok/s
Llama-3.1-8B-Instruct-4bit on M3 Ultra130.2 tok/s
Qwen3-8B-4bit on M3 Ultra120.5 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.