WebJev
Open Apache-2.0 browser-agent decision model (Qwen3.5-35B-A3B fine-tune) that doubled Jev 1.13's task success in the same agent.
Overview
WebJev is a Jev-like decision model specialized for browser use. Given a live web page, it picks the next operation and target element in a single forward pass, answering the per-step Choice questions of Browser Use's jev-ultrafast agent behind the same /v1/systemone API. It is a fine-tune of Qwen3.5-35B-A3B, with Apache-2.0 weights, training data and recipe, and a demo app.
On 125 hand-picked real-website tasks graded by deterministic verifiers, with only the decision model swapped, WebJev completed 38.5 percent versus 16.7 percent for Jev 1.13. On eight public single-step decision benchmarks the two were close (83.7 versus 84.7 percent). A showcase of six tasks where WebJev succeeded and Jev failed was re-run three times per model: WebJev succeeded in 16 of 18 takes and Jev in 0 of 18, with failure modes such as giving up at cookie dialogs.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which operation should the browser agent perform next? | — |
| Choice | Which page element should the operation target? | — |
How it works
- The jev-ultrafast agent observes the live page.
- WebJev picks the next operation and target element.
- The agent executes it in the browser.
- A deterministic verifier grades task completion.
Strengths
- Much higher end-to-end success than Jev 1.13 in the same agent.
- Open weights, data and recipe.
- Drop-in via the /v1/systemone API.
Caveats
- Large model (35B total, 3B active) needs serious hardware.
- Tasks were hand-picked; showcase is selected successes.
At a glance
- Type
- Service
- Status
- research
- Language
- Python
- Runtime
- GPU model server (/v1/systemone)
- Has code
- yes
- License
- Apache-2.0
Use cases and integrations
How curated lists describe it
Browser agents: open-weight Apache-2.0 decision model (a Qwen3.5-35B-A3B fine-tune) that answers jev-ultrafast's per-step `Choice` questions for the next operation and its target element behind the same `/v1/systemone` API, completing 38.5% of 125 hand-picked real-website tasks graded by deterministic verifiers versus 16.7% for Jev 1.13 in the same agent.
yibie/awesome-jevJev-compatible decision model for browser agents: a Qwen3.5-35B-A3B fine-tune that picks the next operation and target element in the Jev Ultrafast loop, behind a `/v1/systemone`-compatible server. On 125 real-website tasks graded by deterministic verifiers, it completes 38.5% vs 16.7% for Jev 1.13 in the same agent. Apache-2.0 weights, training data and recipe, and a demo app.
AnotiaWang/awesome-decision-models
Frequently asked questions
How much better is it than Jev for browsing?
38.5 versus 16.7 percent task success on 125 real-website tasks in the same agent.
Is it better at single-step decisions?
No, the two are close (83.7 vs 84.7 percent).
Which agent does it plug into?
Browser Use's jev-ultrafast.
Related projects
Local browser agent that picks each step's operation and element with a Jev-shaped model such as a fine-tuned Laya, no cloud needed.
MCP serverbetaPythonBrowser-agent benchmark with five timed events from one prompt, plus a ~200-line DevTools agent where Jev picks each action.
ProjectbetaJavaScriptComputer-use platform (Cua Spaces, Driver, Lume) with CUA-S1 small decision models and a preview jev-use example for bounded actions.
ProjectbetaPythonego-decision-layer: measured, pluggable decision layer for ego lite where Jev picks operation and target in one request.
ProjectbetaTypeScriptPlugs any CDP browser, including engines that never draw a page, into a Jev loop that picks operations and targets from the DOM.
LibrarybetaPythonBrowser automation where Jev picks every step from a sightmap's named actions and reports whether the goal is met.
ProjectbetaGo