The weights are byte-identical to the upstream export by onnx-community; nothing was re-quantised or fine-tuned here. This repo exists so the extension depends on a copy that cannot be renamed or deleted out from under its installed users.

Only the files the browser runtime loads are included -- the q4f16 graph plus tokenizer and config. For every other precision, use the upstream repo.

Use

import { pipeline } from "@huggingface/transformers";

const pipe = await pipeline("text-generation", "latentdivergence/mole-lfm2-1.2b-q4f16", {
  dtype: "q4f16",
  device: "webgpu",
});

Licensed as per the base model.