kev-0.6b-browser-use-ONNX

The WebGPU build of kev-0.6b-browser-use: merged and 4-bit, 348 MB, loads with open-jev and Transformers.js.

A tiny 0.6B Jev-like model, fine-tuned from Kev-0.6B for browser use. It follows the graph layout of onnx-community/kev-0.6b-ONNX.

Usage

npm install open-jev @huggingface/transformers

import { OpenJev, choice } from "open-jev";

const kev = await OpenJev.load({ model: "arbazsiddiqui/kev-0.6b-browser-use-ONNX", dtype: "q4f16" });

const state = `Goal: Type "Alan Turing" into the search box
Candidate elements:
[0] <a> role=None "Main page"
[1] <input> role=searchbox "Search Wikipedia"
[2] <button> role=None "Search"`;

const { element, operation } = await kev.decide(state, {
  element: choice("Which element should be acted on?", [
    '[0] <a> "Main page"',
    '[1] <input> "Search Wikipedia"',
    '[2] <button> "Search"',
  ]),
  operation: choice("What operation should be performed on the target element?", ["CLICK", "TYPE", "SELECT"]),
});

element.choice;     // '[1] <input> "Search Wikipedia"', confidence 0.93
operation.choice;   // "TYPE", confidence 0.74

Without open-jev: the graph takes input_ids and attention_mask built with Kev's five delimiter tokens (listed under kev in config.json) and returns one logit per token. Read the logit at each option's closing delimiter and softmax within each question. open-jev's source is the reference encoder.

Files

file	size
onnx/model_q4f16.onnx + onnx/model_q4f16.onnx_data	348 MB

Provenance

• Base: Qwen/Qwen3-0.6B-Base, LoRA r=16 merged, pointer head dim 256.
• Fine-tune: arbazsiddiqui/kev-0.6b-browser-use, a Kev-0.6B fine-tune on Mind2Web plus WebChain browser-action data, trained with jaredpalmer/kev's own trainer (github.com/jaredpalmer/kev).
• This repo: ONNX export only, built with train/export_kev_onnx.py in github.com/arbazsiddiqui/kev-browser-use, following the graph layout of onnx-community/kev-0.6b-ONNX.

License

Apache-2.0 for the adapter and head. The base model is Apache-2.0 (Qwen3).