Is a 0.57 mean score against the teacher strong, or a flattering metric?

Flattering in places. It is (model - random) / (teacher - random), averaged over ten games with 16 held-out episodes each and argmax move. Three games reach 1.00 but Breakout is 0.14, Floppy Bird 0.18 and 2048 0.21, and where the teacher sits far above random the ratio compresses the gap.

Will it run on my laptop, or do I need datacenter hardware?

You need a CUDA GPU. The README quotes about 3 GB for inference and 17 GB for training at batch 64, plus Python 3.12 and playwright install chromium. The 43 ms per move is measured on an H200; no consumer-GPU latency is documented.

Does it remember anything between moves, or is every decision blind?

Each decision is one still frame. The vision tower's temporal patch of 2 could carry a second frame at no extra token cost and this release leaves it off, and the model never sees a game's name, only the current frame and a shuffled list of moves.

So where is the Jev part here, is this just Qwen with a fine-tune?

The contract, not the weights. Qwen3.5-0.8B-Base is fine-tuned locally on 2.2M frames and served at /v1/systemone in the OpenJev request shape, returning a probability per move in one forward pass. No TypeSafe service is involved. General ability held up: a fifth of each epoch went to image and text questions, beating the base on MMBench and matching it on MMLU.

What ten games does it play?

Tetris, Snake, Pacman, Javascript Racer, Space Invaders, Sokoban, Infinite Mario, Floppy Bird, Breakout, and 2048. Each tile opens that game on the demo with the model playing from raw pixels only.

How do I run one game from the command line?

After setup, run python -m playjev.play snake --policy local --ckpt OmniJev/PlayJev-0.8B --episodes 1. It pulls the weights from Hugging Face, opens Snake in headless Chromium, and plays it.

How do I serve the model as a System One endpoint?

python -m playjev.serve --ckpt --port 18732 serves the checkpoint at /v1/systemone in the OpenJev request shape. The demo switches every tile to it with the ?server=http://127.0.0.1:18732 query parameter.

Jev directory

CommunityJev Repos & SDKs

PlayJev — OmniJev

Qwen3.5-0.8B fine-tuned to play ten classic browser games from raw pixels: one frame in, one forward pass, one move out in 43 ms on an H200. Public weights and a browser demo make it an open, vision-only take on Jev-shaped game decisions.

See highlights Open github.com

Stars: 55
Language: JavaScript
License: APACHE-2.0
Reading time: 3 min
Published by: Community
Author: OmniJev
Last verified: Sep 2026

OmniJev/PlayJev

Organization repository on GitHub

View on GitHub

🚀🚀 A 0.8B JEV-like multimodal model playing GUI games directly from raw pixels.

Forks: 5
Last pushed: Updated Sep 2026

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

About this project

PlayJev: A Multimodal JEV-Like Model for Small Games. 🎮   Show the live demo   🎮. PlayJev is Qwen3.5-0.8B-Base fine-tuned to play ten classic browser games from raw pixels. One frame goes in, one forward pass runs, one move comes out, 43 ms on an H200. Every picture above is the trained model playing, each a frame from a recorded held-out episode with the score it had reached by then. The weights are on Hugging Face, with every record we trained on beside them in data/.

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

Tagscommunity javascript games open-models demo

Works forJev for Scoring

Used inJev for Gaming

Highlights

Qwen3.5-0.8B fine-tuned on 2.2M frames to play ten browser games from raw pixels.
One frame in, one forward pass, one move out in 43 ms on an H200.
Returns a probability per move; moves are shuffled in training so position carries no information.
Weights are on Hugging Face, and every game in the demo is playable in the browser.
Ten games share one window.pj hook: start, step, frame, score, done, actions.

Quickstart

bash

pip install -r requirements.txt
playwright install chromium

Watch out

Apache-2.0. Bundled game art and levels keep their source licenses (Nintendo sprites, Microban levels), the released weights see a single still frame per decision, and running it needs pip install -r requirements.txt plus playwright install chromium.

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.

Hacker News: Show HN: OpenDecision, a local decision model with a Doom demo.

dwa35926 pointsShow HN: OpenDecision – a 400M zero-shot model makes local decisions, plays DoomRead the thread on Hacker News

X: Standard One SH (StandardThinking)

LinkedIn: What multimodal decision models actually change, and what they do not.

LinkedInJevDirectory.org

Hacker News: Decision Index 0.2.1 scores open reproductions of Jev on a frozen suite.

jackbravo4 pointsDecision Index 0.2.1 – 70 open reproductions of TypeSafe Jev's Decision ModelRead the thread on Hacker News

Other listings of this project

Forks and copies of this project in the directory. This entry is the canonical one.

PlayJev

A browser visual IDE for composing Jev API requests: build a state plus Choice, Score, Noul and Box questions with image assets, preview and validate the request JSON, send it, and read the answers as visual cards instead of raw JSON.

Jev Demos & Experiments#community#github#typescript

Community

MoeclubM/PlayJevTypeScript

playjev

TypeScript library adding plain-English browser control to an existing Playwright page, where Jev picks numbered nodes and fixed operations and Playwright executes them.

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

Community

filedcom/playjevTypeScript

From the Jev guides

Original write-ups on what this entry is for.

The Open Reimplementations: Local Models That Speak System One

Sep 22, 2026

Closed weights have not stopped the ecosystem from rebuilding the shape in the open. API-compatible servers, retrained heads on small models, and inference-time tricks that skip generation entirely.

Jev Repos & SDKs#open-models#research#local

Read article

From the Jev AI model community

Posts from builders shipping with Jev right now.

Follow @typesafeai