wev: Distilling LLM Browser Agents into
Open, Local System-One Decision Models

Jun Huang^* · Xin Ren^*

University of Electronic Science and Technology of China · ^*Equal contribution

📄 Paper 💻 Code 🤗 Models 🗂️ Dataset pip install wev-ai

A local decision model for agents: typed questions in, calibrated probabilities out, in one forward pass. No text generation, no API key.

Interface distillation: a teacher LLM answers a browser agent's typed requests on live websites, an LLM judge keeps the verified episodes, and the wev student serves the same interface locally.

76%

browser steps right on unseen websites (next best open model: 21%)

30 / 153

live-website tasks completed, on par with the qwen3-max teacher (27)

10–320 ms

per decision on one RTX 5090; 77 ms on a laptop

36%

of the teacher's own DONE claims rejected by the judge

Abstract

System-One decision models answer typed questions about a free-form state with a probability for every declared option, in one forward pass and without generating text, but the open models of this kind are trained on general decision data and reach at most 21% step success on browser steps from unseen websites. We observe that the typed interface between a browser agent and its System One is itself a source of supervision: a prompted LLM placed behind it while the agent works on live websites produces training data in exactly the decision model's format, and an LLM judge reading each episode's final page decides which episodes to keep. We combine this interface distillation with Mind2Web and NNetNav demonstrations converted to the same format, after auditing NNetNav's stopping labels, and with general decision corpora. The resulting open models, wev-1.7b, wev-4b and wev-8b, build on Qwen3 base models. wev-4b answers 76% of browser steps on unseen websites correctly and stays competitive on general decision benchmarks. As the System One of an open browser agent, it completes as many held-out live-website tasks as its LLM teacher, at 15–320 ms per decision on one consumer GPU.

Results

Browser step success against out-of-domain general accuracy

Only wev is strong on both browser steps and general decisions.

Model	Browser step	Transfer	Typed-dec.
wev-4b	75.9	73.8	79.4
wev-8b	75.5	77.2	79.1
wev-1.7b	68.2	65.5	79.5
Kev-4B	21.2	82.1	65.1
Kev-8B	19.0	76.8	62.7
Laya (typed-dec.)	0.7	62.8	76.8
Laya	0.0	63.7	36.2

Test splits, same scorer for every model. Browser step: Mind2Web step success on unseen websites. Transfer: Kev transfer-v4, out of domain for all models. Typed-dec.: typed-decisions, on which wev and Laya (typed-dec.) train.

Models

Model	Size	Best for	General decision	Browser step
wev-4b	8.1 GB	default; best on live websites	15 ms	217 ms
wev-8b	15.2 GB	most accurate out of domain	21 ms	322 ms
wev-1.7b	3.5 GB	laptops; the fastest	10 ms	91 ms

Median latency on one RTX 5090 (bf16), one request at a time.

Quickstart

pip install "wev-ai[serve]"

import wev
m = wev.load("alanhuangya/wev-4b")
out = m.predict(
    state="Refund request: order #4411 arrived damaged, customer attached photos.",
    questions={"action": {"type": "choice", "instructions": "What should support do?",
                          "criteria": {"refund": "Refund the order.", "replace": "Ship a replacement.",
                                       "escalate": "Send to a human agent."}}},
)

# or serve a drop-in POST /v1/systemone endpoint
wev serve --model alanhuangya/wev-4b --port 8009

Citation

@misc{huang2026wev,
  title     = {wev: Distilling LLM Browser Agents into Open, Local System-One Decision Models},
  author    = {Huang, Jun and Ren, Xin},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.22941164},
  url       = {https://doi.org/10.5281/zenodo.22941164}
}

wev: Distilling LLM Browser Agents intoOpen, Local System-One Decision Models