tok/s arena · your box vs their box
Fastest tok/s on that setup.
Same hardware class. Same model. Different engines, quants, flags — each is its own event. Ranked by personal best signed tok/s. Your agent signs on your machine. We do not certify it.
Pick a class and a model below — that is the event. The generate runs on your Ollama at 127.0.0.1. If the tag is not pulled, the bench dies. We do not download weights for you.
how to run it get a token
5090
4090
3090
A6000
L40S
Spark
2× Spark
2× 4090
2× 3090
H100
Strix Halo
M3 512
qwen3.8-27b
qwen3.8-27b weights
qwen3.6-35b-a3b
qwen3.6-35b-a3b weights
gemma-3-27b
gemma-3-27b weights
mistral-small-3.2-24b
mistral-small-3.2-24b weights
llama-4-scout-17b-16e
llama-4-scout-17b-16e weights
lone · n=1 signed · agreement —
qwen3.8-27b ollama q4_k_m num_predict=256,temperature=0 ctx 4096
median 131.4 tok/s · range 131.4–131.4
| # | handle | PB tok/s | runs | flags | host |
| 1 |
test |
131.45 |
1 |
num_predict=256 temperature=0 |
|