Intelligence vs. Speed on your GPU
Local Intelligence Index (higher is better) vs. decode tokens/s (higher is better) · best engine per point ·
● filled = fits in VRAM · ○ hollow = CPU/RAM offload · size = file size · dashed = Pareto frontierillustrative data
Decode speed by GPU
Output tokens per second · higher is better ·
Missing bar = engine not supported or model does not fit with that engineillustrative data
What should I run?
Best quality that meets your speed floor · 64 GB system RAM
Soundings
All runs for the selected GPU · click a column to sort
| Model | Quant | Engine | File | Decode t/s | Prefill t/s | Index | Mode | Status |
|---|
Status: ● verified = our run or ≥ 2 matching community runs · ● community = single submissionillustrative data
Ping
One standard run for every model × quant × engine × GPU: same prompts, contexts 512 / 8k / 32k, fixed seeds, warm-up excluded.
Echo
We record prefill and decode speed, TTFT, peak VRAM/RAM and KLD vs. BF16. Raw logs and the exact engine commit ship with every run.
Chart
Run it yourself: fathometer run <model> --engine llama.cpp then fathometer submit. Coming soon.