TL;DR of the article, generated automatically.

Kiln is a browser library that runs AI models on WebGPU, straight from original weights.

• Speed: 2.2× median faster than transformers.js across 31 task and model pairs, per the repo.
• Benchmarks: measured on a 2017 Intel Mac with integrated graphics, in Chrome.
• Slower cases: cached model loads, 4-bit download size, and full masked-LM logits.
• Loading: reads Hugging Face or GGUF files directly, no ONNX export. MIT licensed.