phi-4

on RTX 5090

141

tok/sdecode

RTX 5090 (32GB) + AMD Ryzen 7 9850X3D 8-Core Processor (8c) + 30GB

rank in tier

41/165RTX 5090 runstop 25%

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

Qwen2.5-7B-Instruct-4bit on M3 Ultra140 tok/s
Qwen2.5-Coder-7B-Instruct-4bit on M3 Ultra139 tok/s
Llama-3.1-8B-Instruct-4bit on M3 Ultra130 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.