wev-4b

Jun Huang*, Xin Ren* · University of Electronic Science and Technology of China · *Equal contribution

A local decision model: typed questions in, calibrated probabilities out, in one forward pass. wev-4b answers the POST /v1/systemone request shape (choice, yes/no and score questions over a free-form state), for general decisions and for browser-agent steps (which operation? which element?). It runs on your own GPU: no API key, no per-call cost, nothing generated.

Code, training and evaluation: alanhuangyoo/wev. Paper: doi:10.5281/zenodo.22941164. Independent project; not affiliated with TypeSafe AI.

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, first refund this year.",
    questions={
        "action": {"type": "choice", "instructions": "What should support do?",
                   "criteria": {"refund": "Refund the order.", "replace": "Ship a replacement.",
                                "escalate": "Send to a human agent."}},
        "fraud_risk": {"type": "noul", "instructions": "This request looks fraudulent.",
                       "criteria": {"true": "Likely fraud.", "false": "No sign of fraud."}},
    },
)
print(out["answers"])

wev serve --model alanhuangya/wev-4b --port 8009   # drop-in POST /v1/systemone, e.g. for jev-ultrafast

Results

Test splits, held out from training; every other model was run on the same requests and scored the same way (per-question accuracy; scripts/compare.py). For wev-4b and wev-8b, two candidates each were read on test (see the repository README).

General typed decisions

model	kev decision-v7	kev transfer-v4	typed-decisions
wev-4b	80.6	73.8	79.4
Kev-4B	88.2	82.1	65.1
Kev-8B	88.1	76.8	62.7
Laya (typed-decisions)	65.7	62.8	76.8
Laya	64.3	63.7	36.2

wev-4b trains on 80% of the typed-decisions train split, like the Laya (typed-decisions) specialist; Kev and Laya do not, so on that column they are generalists. kev decision-v7 is Kev's own training suite (wev-4b also trains on its train split); transfer-v4 is out-of-domain for every model here.

Browser steps (Mind2Web test split: websites unseen in training, jev-ultrafast request format; step success = operation and target element both right)

model	step success	operation
wev-4b	75.9	91.2
Kev-4B	21.2	35.7
Kev-8B	19.0	73.3
Laya (typed-decisions)	0.7	13.1
Laya	0.0	2.5

873 requests; 11 exceed the context wev-4b is evaluated with and count as wrong for it.

NNetNav test split (live-web steps, DONE judged by an LLM): step success 61.4, DONE recall 84.5, premature DONE 10.0.

End to end (153 held-out tasks on live websites, run by jev-ultrafast with wev-4b as its System One; success = the agent says DONE and an LLM judge reading the final page agrees): 30/153 (19.6%) tasks, vs 27/153 (17.6%) for the qwen3-max teacher behind the same agent. Live sites differ from run to run; treat gaps of a few tasks as noise.

Model

• Backbone: Qwen/Qwen3-4B-Base without its vocabulary head, LoRA r=16 on every attention and MLP projection, merged into the weights of this export; 36 layers, bf16.
• Readout: a pointer head scores each option's </opt> state against the question's <decide> state.
• Each question sees the state and itself only (block-causal branches, positions restart after the state), so a request with many questions costs one pass and answers never depend on question order.
• Context: state up to 4096 tokens, each question up to 8192 tokens (trained with 2048); longer page states are shrunk before encoding.

Training

1 epoch, lr 0.0001, one-cycle schedule, soft-label cross-entropy where the source has soft labels. Recipe and data builders: alanhuangyoo/wev.

source	license	what it adds
Mind2Web	CC BY 4.0	human browser steps: click, type, select
NNetNav-live	Apache-2.0	live-web steps; DONE relabelled by an LLM judge
teacher episodes	outputs of qwen3-max	jev-ultrafast on live sites with qwen3-max as System One, success judge-verified
kev decision-v7	per source	ten public classification / QA sources plus rule records
typed-decisions	Apache-2.0	agent / ops workflows, 5 questions per case (80% of train)
tasksource-jev	mixed (per source task; some research-only)	hundreds of classification tasks as decisions
jev-distill-corpus-v3	Apache-2.0	synthetic operational scenarios, soft labels
typed-decisions-synth	MIT	multi-question cases over 149 domains

Use terms. Some training data carries its own terms: several tasksource-jev source tasks are research-only, and the teacher episodes are qwen3-max outputs subject to its provider's terms. Treat this model as a research artifact and check those terms before any commercial use.

Limitations

• English only. Decisions, not text: TYPE values come from a separate text model, as in jev-ultrafast.
• Browser targets are scored among the candidates the agent lists (8–40 per step), not every element on the page.
• DONE and BLOCKED are the hardest operations; gate DONE on its probability when early stops are costly.
• Not compared with Jev itself (no API access).

Citation

Jun Huang and Xin Ren contributed equally (University of Electronic Science and Technology of China).

@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}
}

License

Apache-2.0, like the base model. Architecture code adapted from kev (Apache-2.0).