kleinhirn device benchmark
Measure how kleinhirn runs on this device: same answers as the PyTorch reference, and how fast. The model is GLiNER2.5-small, the tests are 128-token and 256-token inputs. It takes a few minutes. Downloads: f16 about 152 MB, f32 about 297 MB (shared with wasm), all stages about 449 MB.
Privacy
Nothing is uploaded. The page talks to its own origin and to the weights host, nothing else. There are no scripts, fonts, analytics or cookies from third parties. The result file contains WebGPU adapter info and limits, which can roughly identify a device, so read it before you share it.
Before you start
- Plug the device in.
- Keep the screen on and this tab in front. Background tabs are throttled.
- Turn low power mode off.
- Close other tabs and heavy apps.
This device
Collecting.
Stages
Fallback chain, in order. The page first loads the engine the way an app would (auto) and reports which stage it picks, then measures each selected stage on its own.
Result
To contribute, open a GitHub issue with the device-result form and drag the downloaded file into it. Parity is checked again by script. Latency is what your device reported.