A full-fledged agent for programming on a model with only 2 billion parameters was launched in the browser. The Pi runs entirely inside the browser via WebGPU. It is based on MiniCPM5-2B, Transformers.js and ONNX four-bit weights. All previous attempts to create such an agent had failed, but MiniCPM5 was powerful enough to make it suitable for real work. The model downloads once and takes up about 1.84 GB. The agent can then work with your files directly in the browser. You can try it now on Hugging Face. 🤖