Best open-weights model that runs in a browser
kev-0.6b-browser-use
A decision model for browser agents. It reads a goal and the page's candidate elements, then returns which element to act on and whether to click, type or select, as probabilities, in one forward pass. It never writes text, so every answer is an element already on the page.
- Result
- 32.2% step success on 1,000 Mind2Web test steps, from 14.1% for Kev-0.6B as released
- Size
- 0.6B parameters · 348 MB as a 4-bit build · 80 ms a decision on an NVIDIA L4
- Training
- Mind2Web's training split (13.8K steps), then WebChain for four times the rows
- Runs on
- WebGPU in the browser through open-jev, or Kev's own server with the LoRA adapter
- Links
- Live demo · Hugging Face · GitHub
The demo
The model plays A Dark Room, the text adventure, running unmodified on the page. Give it a goal like "craft a rucksack" and it picks the button out of everything the game shows. When the game says there is not enough leather, the page asks it how to get more, and it puts villagers on the tannery. Nothing is scripted per goal.
How it compares
On the same 1,000 steps it is level with Kev-9B (32.5) at a fifteenth of the size; the gap is within noise. Bespoke Nimble-9B and decider-2B score 26.2 and 24.7. The hosted Jev API still leads at 43.5, and Mind2Web's own results for bigger fine-tuned models suggest the gap is size.
The full write-up has the results table, the training data and the code to use it.