Builder: KLEINHIRN
The name is German for cerebellum, the small part of the brain that runs the reflexes. kleinhirn runs encoder models in your browser, on your own GPU, with no server in between: classification, embeddings, named entities and reranking, loaded straight from Hugging Face checkpoints across six model families. The GPU kernels are my own, written in WGSL for WebGPU, with a WASM fallback for machines without it. Your text never leaves the device. In every benchmark so far it beats ONNX Runtime Web, Microsoft’s standard for the job, by up to three times, and the engine is 47 KB gzipped.
- engine
- 47 KB gzip
- vs ORT Web
- up to 3x faster
- code
- GitHub