Is WebJev a drop-in replacement for Jev in browser agents?

The README says it follows the same request and response format and serves /v1/systemone, making it a drop-in replacement. On the 125 real-website tasks it reached 38.52% success versus Jev 1.13's 16.67% with only the decision model swapped.

What hardware does WebJev need to run?

The serving instructions target one 80 GB GPU because the base model is a 34.7B-parameter mixture-of-experts. Browser runs can use a local Chrome started with --remote-debugging-port=9222 or the Lexmount cloud browser.

What agent does WebJev drive, and what does it decide?

It runs with Browser Use's jev-ultrafast agent. From live page state (URL, title, visible text, indexed elements and recent actions) it answers one decision request per step, choosing the operation (CLICK, TYPE_TEXT, SELECT, page controls, DONE, BLOCKED) and the target element.

How much better is WebJev than Jev on real websites?

On 125 hand-picked real-website tasks graded by deterministic verifiers, WebJev-35B-A3B completed 38.52% against Jev 1.13's 16.67% in the same agent. On 8 public single-step decision benchmarks the two were close, at 83.72% versus 84.70% mean accuracy.

Does WebJev generate any text?

No. It is a decision model that reads a state and typed questions and returns a probability for every option in one forward pass. A separate text helper model writes the text and the final answer to the user; the decision model only chooses.

Which browsers can I run WebJev with?

Either Lexmount Browser, the recommended cloud option needing LEXMOUNT_API_KEY and LEXMOUNT_PROJECT_ID, or your local Chrome started with --remote-debugging-port=9222 and BROWSER=local. All recorded evaluations and GIFs used Lexmount Browser.

What comes in the WebJev repository besides the weights?

Training code and how the data mixture was built, a vLLM serving stack behind a Jev-compatible API, 8 public single-step decision benchmarks, the 125-task real-website harness with deterministic verifiers, and a browser-agent demo app.

CommunityJev Models & Reimplementations

WebJev

Apache-2.0 decision models for browser agents: a Qwen3.5-35B-A3B fine-tune that answers jev-ultrafast's per-step operation and target-element Choice questions in one forward pass behind a Jev-compatible /v1/systemone API. Includes training code, a serving stack and a 125-task real-website evaluation.

See highlights

Open github.com

Stars: 3
Language: Python
License: APACHE-2.0
Published by: Community
Last verified: Oct 2026

lexmount/WebJev

Organization repository on GitHub

View on GitHub

Jev-like decision models specialized for browser use: given a live web page, they pick the next action in a single forward pass. Includes training, evaluation, and real-website task runs.

Forks: 1
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

Jev-like decision models specialized for browser use. Given a live web page, WebJev picks the next action in a single forward pass. Same browser agent, same live website; only the decision model differs. Left: Jev 1.13. Right: WebJev-35B-A3B. Real time. Verdicts come from the task's deterministic verifier.

Quoted from the project's README. Read the full README.

Tagscommunity browser open-weights decision-models

Works forJev for Decision gating Jev for Tool calling Jev for Classification

Used inJev for Software & DevTools

Languages: Python
Platforms: Browser
Capabilities: Open sourceSelf-hosted

Highlights

WebJev-35B-A3B is a fine-tune of Qwen3.5-35B-A3B-Base: 34.7B total parameters, 3B active per token (MoE), BF16 weights on Hugging Face.
Completed 38.52% of 125 hand-picked real-website tasks graded by deterministic verifiers, versus 16.67% for Jev 1.13 in the same agent.
Single forward pass answers operation, click target, type_text target and select target questions from live page state.
Mean accuracy across 8 public single-step benchmarks was 83.72%, close to Jev 1.13's 84.70%.
Repository includes training code, a vLLM server behind a Jev-compatible API, and the 125-task evaluation harness.

Watch out

Apache-2.0 code and weights, independent of TypeSafe. Serving the 34.7B MoE needs one 80 GB GPU, and the recorded evaluations used a Lexmount cloud browser (local Chrome is also supported).

Questions about this entry

Short answers drawn from the linked project or page.

Reactions & coverage

Posts, threads, and videos about this entry from around the web.

X: Jev driving real Chrome from Grok Bot

X: jev-browser, a general browser skill

X: A second-hand shopping agent

X: A job crawler that reads careers pages with Jev

More Jev resources like this

OpenDecider Nano

A 400M-parameter Jev-compatible decision model with a browser build: 0.796 on typed decisions against Laya's 0.766, 17 ms per question on an L40S, and an ONNX file that runs in a tab.

Jev Models & Reimplementations#community#open-models#local

Community

RSI-Jev

109

Open decision models from a self-improvement loop that publishes every experiment, failures included. They answer Noul, Choice and Score questions on the same POST /v1/systemone wire format as Jev, returning a calibrated probability per option in one forward pass; the 27B scores 65.64 on the Decision Index 0.3 set.

Jev Models & Reimplementations#community#decision-models#system-one

Community

Shanghua-Gao/RSI-JevHTML

Visual-Jev

A Jev-style typed-decision baseline built on the frozen Qwen3.5-4B checkpoint that accepts a local image path directly, letting the vision-language model score declared options without generating text. Ships scoring modes, a WebGPU browser demo and benchmark results measured on an RTX 3090.

Jev Models & Reimplementations#community#multimodal#open-models

Community

jiangxiluning/Visual-JevPython

Jebadiah

Open System One decision models (27B, 9B and 4B on Qwen chat bases, in bf16, GGUF and MLX) that answer Choice, Noul and Score questions with a probability for every option from one forward pass. The repo carries the trainer, data builders, evaluation harnesses and a record of every run.

Jev Models & Reimplementations#community#decision-models#open-weights

Community

getainode/jebadiahPython

jevos

1.3k

A local yes/no decision model that speaks Jev's wire format on a laptop CPU, refusing Choice and Score until those question types ship.

Jev Models & Reimplementations#community#python#open-models

Community

feder-cr/jevC++

AutoJev

130

An open recipe for a 27B multimodal decision model trained with full-weight SFT on 73,000 examples over one H200, serving choice, noul and score on /v1/systemone with per-checkpoint provenance and released calibration plots.

Jev Models & Reimplementations#community#calibration#qwen

Community

denis-pplx/autojevPython

From the Jev guides

Original write-ups on what this entry is for.

Multimodal Decision Models: What Clef and d1 Change

Oct 12, 2026

Jev takes text only; Clef-omni adds image, audio, and video input and Liquid d1 adds images, each keeping the same typed-answer shape. Where that matters, where it does not, and the costs the vendors underplay.

Jev Models & Reimplementations#multimodal#open-models#decision-model

Read article

The Decision-Model Price War

Oct 11, 2026

One model in September, a dozen by October. Clef, Liquid d1, Perplexity, and OpenAI all ship the same typed-answer shape now. How to read the rival benchmarks and pick on your own task.

Jev Models & Reimplementations#decision-model#benchmarks#pricing

Read article

What Is Actually on Hugging Face

Oct 8, 2026

Hundreds of repositories carry the system-one tag, but the ecosystem is a dozen ideas in different sizes. A tour of the families, the licenses, and what a download does not include.

Jev Models & Reimplementations#open-models#hugging-face#local

Read article

Back to all resources

From the Jev community

Posts from builders shipping with Jev right now.

Follow @typesafeai