Are the benchmark numbers real, and what exactly does the 76% measure?

It is the Mind2Web test split, 873 requests on websites unseen in training, where a step counts only when both the operation and the target element are right. Every model in the table receives the same requests and is scored the same way, taking each model's most probable option as its answer.

How is this different from the hosted Jev endpoint I could just call?

Same request and response shapes, so an agent written for a System One API can point at 127.0.0.1:8009 instead of the hosted endpoint. you supply the GPU and there is no per-call billing. Note that jev-ultrafast hard-codes that URL in one file, so integration is a one-line change.

Can I run this on a laptop without a discrete GPU?

Yes, via wev-1.7b, a 3.5 GB download quoted at 10 ms median for a general decision on an RTX 5090, with 77 ms cited for a laptop. Calibration and out-of-domain accuracy both fall off relative to the 15.2 GB 8B model, which is the one described as most accurate out of domain.

What does the training data look like, and is the DONE label trustworthy?

Episodes come from a prompted LLM answering the typed interface while an agent browses live sites, kept only when an LLM judge confirms the final page. That judge rejected 36% of the teacher's own DONE claims, and a separate audit overturned 31% of NNetNav's stopping labels. Mind2Web and NNetNav steps are converted to the same request format.

How do I install and serve OpenJev?

pip install "openjev-ai[serve]" installs the package imported as wev; wev.load("alanhuangya/OpenJev-4B") runs it in Python, and wev serve --model alanhuangya/OpenJev-4B --port 8009 exposes POST /v1/systemone, the same shape a System One API expects.

What license is it and where do the weights live?

Apache-2.0, with models OpenJev-1.7B, 4B, and 8B on the Hugging Face Hub plus an OpenJev-Data dataset. OpenJev-4B v0.2 continues training from the Kev-4B checkpoint, which is also Apache-2.0.

How do I point a browser agent at it?

An agent written for a System One API, such as jev-ultrafast, can call http://127.0.0.1:8009/v1/systemone instead of the hosted endpoint. jev-ultrafast hard-codes that URL in jev_ultrafast/model.py, so change that one line.

CommunityJev Demos & Experiments

Wev

Local System One decision models in 1.7B, 4B and 8B sizes that answer choice, yes/no and score questions over the POST /v1/systemone shape, with no text generation and no API key.

See highlights

Open github.com

Stars: 60
Language: Python
License: APACHE-2.0
Published by: Community
Author: alanhuangyoo
Last verified: Oct 2026

alanhuangyoo/wev

User repository on GitHub

View on GitHub

OpenJev: open-source, local alternative to Jev. System-One decision models: typed questions in, calibrated probabilities out; general decisions and browser-agent steps.

Forks: 0
Last pushed: Updated Oct 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

The open-source, local alternative to Jev: System-One decision models for agents. Typed questions in, calibrated probabilities out, in one forward pass. No text generation, no API key. Python package wev-ai (import wev) · models OpenJev-1.7B / 4B / 8B on the Hugging Face Hub. Quickstart · Results · How it works · Train your own · Citation.

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

Tagscommunity awesomejev catalog-import

Works forJev for Classification Jev for Tool calling

Used inJev for Software & DevTools

Languages: Python

Highlights

Mind2Web step success of 75.9% for wev-4b on sites unseen in training, against 21.2% for Kev-4B; operation-only 91.2%.
Median latency on one RTX 5090: 10-21 ms for a general decision and 91-322 ms for a full browser page.
As the System One of jev-ultrafast it completed 30 of 153 held-out live tasks, against 27 for the teacher LLM it was distilled from.
Trained by interface distillation, with an LLM judge rejecting 36% of that teacher's own DONE claims.

Watch out

Figures are the author's, and test splits were held out except that two candidates each for wev-4b and wev-8b were both read on test. wev is evaluated with a 4,096-token state, so the 11 Mind2Web requests beyond it count as wrong. Weights download from the Hugging Face Hub. Copy reviewed against the repository README on 2026-10-04, not from running it.

Questions about this entry

Short answers drawn from the linked project or page.

More Jev resources like this

jev-kit

Claude Code kit that guards PreToolUse calls with a free pre-filter plus one Jev question for ambiguous cases, and adds sub-agent sizing, rechecks, search, daemon speedups, and monitoring.

Jev Demos & Experiments#community#awesomejev#catalog-import

Community

jonathanavis96/jev-kitPython

jev-ultrafast-mcp

MCP server where one browser_goal call hands a URL, a goal, and code-checked assertions to a server-side loop driven by a decision model over live page elements.

Jev Demos & Experiments#community#jevbest#catalog-import

Community

jiawei686/jev-ultrafast-mcpPython

aside-jev

Browser workflow that keeps one Aside REPL session open, compacts page observations into allowed candidates, and has Jev pick the next action with confidence checks.

Jev Demos & Experiments#community#awesomejev#catalog-import

Community

himomohi/aside-jevPython

keys-mcode-continuous-context-browser-decision-enhancement-pack

Enhancement pack for the MiniMax Code CLI with TUI patches, continuous context renewal, a local-vLLM compaction fix, and a measured arm64 browser lane.

Jev Demos & Experiments#community#awesomejev#catalog-import

Community

drowzeys/keys-mcode-continuous-context-browser-decision-enhancement-packPython

jev-harness

Decision layer for coding agents that triages test failures deterministically before spending frontier tokens. Offline heuristics handle missing packages, flaky transients, and loop breaking, with optional Jev micro-decisions and CLI, MCP, hook, and SDK use.

Jev Demos & Experiments#community#awesomejev#catalog-import

Community

ismaelsoilet/jev-harnessPython

open-spark-jev

Independent open-source local decision model inspired by Jev, with a spark-s1 family plus smaller control models for agent gates, served locally with probabilities and policy actions.

Jev Demos & Experiments#community#jevbest#catalog-import

Community

abhishek085/open-spark-jevPython

From the Jev guides

Original write-ups on what this entry is for.

Per-Tick Decisions in Games

Sep 28, 2026

A game tick is a state plus a legal action list. How Clash Royale, Minecraft, and the earlier Mario, Pokémon, StarCraft, and drone loops ask Jev one step at a time.

Jev Demos & Experiments#gaming#agents#control

Read article

Back to all resources

From the Jev community

Posts from builders shipping with Jev right now.

Follow @typesafeai