Trust / Backends
Backends
A backend is the software that runs the model on the laptop. The same model file can behave differently on two backends, so every run is labelled with the backend that produced it.
Speeds compare only within a backend. A tokens-per-second figure measured on one backend says nothing reliable about the same model on another, so the speed pages and the explorer should be read one backend at a time.
Runs by backend
ROCm 7.1
9 published runs, 216 published model results; Ollama 0.34.4, 0.40.13, 0.40.4.
| Run | Date | Setup recorded with the run | Models |
|---|---|---|---|
r20261105-bv1-2abe78 |
2026-11-14 | ROCm 7.1 (Ollama 0.40.13, flash attention, 16k context; 8 parallel slots) | 9 |
r20261102-bv1-5246ab |
2026-11-05 | ROCm 7.1 (Ollama 0.40.13, flash attention, 16k context; 8 parallel slots) | 33 |
r20261029-bv1-3e4075 |
2026-10-29 | ROCm 7.1 (Ollama 0.40.4, flash attention, 16k context; 8 parallel slots) | 1 |
r20261028-bv1-ef0895 |
2026-10-28 | ROCm 7.1 (Ollama 0.40.4, flash attention, 16k context; 8 parallel slots) | 2 |
r20261028-bv1-7ce712 |
2026-10-28 | ROCm 7.1 (Ollama 0.40.4, flash attention, 16k context; 8 parallel slots) | 4 |
r20261028-bv1-2e24cd |
2026-10-28 | ROCm 7.1 (Ollama 0.40.4, flash attention, 16k context; 8 parallel slots) | 1 |
r20261016-bv1-473145 |
2026-10-26 | ROCm 7.1 (Ollama 0.40.4, flash attention, 16k context; 8 parallel slots) | 102 |
r20261022-bv1-54a19b |
2026-10-23 | ROCm 7.1 (Ollama 0.40.4, flash attention, 16k context; 4 parallel slots) | 7 |
r20261006-bv1-53e146 |
2026-10-11 | ROCm 7.1 (Ollama 0.34.4, flash attention, 16k context; 4 parallel slots until 2026-10-07 00:45Z, 8 after) | 57 |
Vulkan
3 published runs, 12 published model results; Ollama 0.34.4, 0.40.4.
| Run | Date | Setup recorded with the run | Models |
|---|---|---|---|
r20261029-bv1-9ddd5f |
2026-10-29 | Vulkan (Ollama 0.40.4, flash attention, 16k context; 8 parallel slots) | 1 |
r20261015-bv1-5ac997 |
2026-10-15 | Vulkan (Ollama 0.40.4, flash attention, 16k context; 2 parallel slots) | 1 |
r20261012-bv1-6a1291 |
2026-10-14 | Vulkan (Ollama 0.34.4, flash attention, 16k context; 4 parallel slots) | 10 |
Models published on more than one backend
No model has a published result on more than one backend yet, so this site cannot yet say how much the backend alone changes a score.
Withheld
These models are left out of every figure on this page. Why each was withheld is on the leaderboard.
- Apertus-8B-Instruct-2509 is withheld on ROCm 7.1; its published result comes from Vulkan.
All withheld models: