What is jeva and what does it do?

A 2B browser-agent decision model you can train and run yourself. It turns a page observation plus a goal into exactly one action as a compact JSON object, such as a click on element 2 or typed text, with no prose.

How does jeva relate to Jev and TypeSafe?

It speaks TypeSafe's System One protocol: jeva systemone serves /v1/systemone, so an existing jev-ultrafast agent is repointed by changing its base URL from api.typesafe.ai to the local server, and nothing else changes.

How fast and large is it?

The README reports about 230 ms per decision running the 1.6 GB Q4_K_M GGUF under llama.cpp on a single V100; the naive Hugging Face path measures 1500 ms. Weights are merged into MiniCPM5-2B, so there is no adapter to pair.

How accurate is jeva on its benchmarks?

It reports 100% task success on its frozen suites, including site3, a held-out site whose label vocabulary and layout never appear in training, and 4 out of 4 correct runs against a live httpbin form it never saw.

How do I install and serve it?

pip install jeva installs the client and CLI with zero dependencies, jeva download fetches the Q4_K_M weights into ~/.cache/jeva with checksums, and jeva serve leaves a server running on 127.0.0.1:8020.

Do its probabilities work as confidence scores?

No. The README says the answers are one-hot choices, not calibrated; in its measurements correct answers carried a median top-1 of 48% while wrong ones carried 80%, so probabilities cannot be used as a confidence signal.

What are the limits when swapping it into an agent?

System One questions never return free text, so a TYPE_TEXT value cannot travel in answers; jeva adds it under an extra jeva key that clients ignoring it simply drop. The README also says its own runs took text values from the agent's text model.

Jev directory

CommunityJev Models & Reimplementations

Jeva — chemany

A self-trainable 2B browser-agent decision model that speaks TypeSafe's /v1/systemone protocol, so an existing jev-ultrafast agent is repointed with a one-line base-URL change. Runs as a 1.6 GB Q4_K_M GGUF under llama.cpp at about 236 ms per decision.

See highlights Open github.com

Stars: 1
Language: Python
Reading time: 2 min
Published by: Community
Author: chemany
Last verified: Oct 2026

chemany/jeva

User repository on GitHub

View on GitHub

A 2B browser-agent decision model. Speaks the TypeSafe System One protocol, so jev-ultrafast is repointed with one line. Answers do not move with the page title or the form layout. 100% on the action suite.

Forks: 0
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.Homepage

This is one of several Jev listings for the same project. The canonical entry is Jeva — liqiiiii.

Tagscommunity github python browser-automation model

Works forJev for Realtime control Jev for Decision gating

Used inJev for Software & DevTools

Languages: Python
Capabilities: Open sourceRuns locally

Highlights

Turns a page observation plus a goal into exactly one action as compact JSON, with no prose.
2B fine-tune of MiniCPM5-2B, merged weights: 1.6 GB as Q4_K_M GGUF, about 230 ms per decision under llama.cpp on one V100.
Speaks TypeSafe's /v1/systemone, so a jev-ultrafast agent is repointed by changing one base URL.
100% on frozen suites including site3, a held-out site whose labels never appear in training.
Zero human labels: 10,686 trajectories from 2,600 tasks, one LoRA pass of about 75 minutes on one V100.

Quickstart

bash

pip install jeva
jeva download
jeva demo

Watch out

README and GitHub metadata reviewed 2026-10-17; the digest truncates the README mid-sentence and the weights themselves live on ModelScope, not GitHub. Accuracy and latency figures are the author's own frozen suites, not independent benchmarks.

Questions about this entry

Short answers drawn from the linked project or page.

From the Jev guides

Original write-ups on what this entry is for.

Beyond text and boxes: multimodal decisions you can run yourself

Oct 16, 2026

Jev-Omni's merged classifier scores text, image, audio, and video options; VideoJev and Laya Vision show what open multimodal decision models accept, what they return, and which numbers are the authors' own.

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

Read article

Jev 1.14: Images, Box, and the Danger Zone

Oct 15, 2026

Jev Day shipped a routine model bump and a deliberate widening of the API: images on both models, a box question type on 1.14 only, and 262k context, all behind danger-zone headers.

Jev Models & Reimplementations#release#vision#danger-zone

Read article

The Open Decision-Model Roundup

Oct 14, 2026

Training recipes, looped architectures, and language ports: what the open side of the typed-decision class shipped this month, and how to pick by runtime and calibration instead of the leaderboard.

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

Read article

From the Jev AI model community

Posts from builders shipping with Jev right now.

Follow @typesafeai

Jeva — chemany

Highlights

Quickstart

Questions about this entry

What is jeva and what does it do?

How does jeva relate to Jev and TypeSafe?

How fast and large is it?

How accurate is jeva on its benchmarks?

How do I install and serve it?

Do its probabilities work as confidence scores?

What are the limits when swapping it into an agent?

More Jev resources like this

From the Jev guides

From the Jev AI model community