WLojek/basal-2.0-mini-onnx-web

basal-2.0-mini: ONNX int4 for the browser (WebGPU)

This is a browser build of Remek/basal-2.0-mini (1.5B, Bielik v3 (Llama architecture)), a Basal decision model by Remigiusz Kinas (rkinas/basal).

A Basal model reads a text and answers a typed question with a calibrated probability per option, in one forward pass.

Files

• model.onnx + model.onnx.data (0.97 GB): ONNX int4 (MatMulNBits) built with the onnxruntime-genai model builder for the WebGPU execution provider. The LM head is pruned to the last token.
• tokenizer.json, tokenizer_config.json, CALIBRATION.json, basal.json: unchanged from the original model.
• basal-web.json: graph inputs and outputs, the model family, and an int4→fp32 yes/no logit correction.

How to use it

Use it with basal-web, part of the open-source Nopescope project, a small JS runtime that follows Basal's prompt format 2 contract, running on onnxruntime-web. It is used by the Nopescope browser extension.

The runtime gives the same answers as the fp32 PyTorch reference on 23 of 23 test prompts. Results and how to reproduce them are in the Nopescope repository.

License

Apache-2.0, the same as the original model. This is a converted, quantised derivative: the weights are quantised to int4 and the graph is exported to ONNX. It is not affiliated with the Basal authors. Please cite and credit the original model.