NanoLLM benchmark
How NanoLLM's presets run on people's hardware: medians over installations, each installation and each machine counted once.
| System | Preset | Generation, tok/s | At depth, tok/s | Prefill, tok/s | Chat, tok/s | MTP speed-up | Installs |
|---|---|---|---|---|---|---|---|
|
unknown
AMD EPYC 7351P
· ubuntu 26 · server · gemma-4-e4b-qat
|
ram-94g |
22.7 | — | 133 | 22.8 | ×1.37 | 1 |
|
unknown
AMD EPYC 7351P
· ubuntu 26 · server · gemma-4-e4b-qat
|
ram-4g |
21.8 | — | 119 | 22.9 | ×1.30 | 1 |
In brackets: p25–p75.
Recent submissions
- 2026-10-20 — unknown · AMD EPYC 7351P · gemma-4-e4b-qat