How is the Laya versus Jev comparison kept fair?

Both sides receive the same state and question shapes through server.py backends at /api/laya and /api/jev, with identical seeds, spawns, and apple sequences, so only the model differs.

Does the slower model just lose because it decides less often?

Pace follows the slower side so latency shows up as decisions per second rather than movement speed, with a warm-up decision, turn budget, passable bodies, and wrapping walls in snake.

What do the models actually decide each tick?

Snake asks steer as a five-way choice plus sprint as a yes/no, while kombat asks a six-way action plus a heavy-attack commit flag, with frame data, blocking, and punish windows in the ring.

Are the benchmark numbers real?

The recorded run had Laya winning snake on speed while Jev answered steering probes more accurately, but pacing, network path, and seeds change results, so export JSON and re-run your own seeds.

How do I run the arena?

Python 3.10 or newer, then clone the repo, pip install -r requirements.txt, copy .env.example to .env with TYPESAFE_API_KEY, and python server.py 8740. The front page is at http://localhost:8740/, with /snake/ and /fight/ for the two arenas. Laya needs no key.

What does the server expose?

server.py serves the page and exposes two backends taking the same state-and-questions payload and returning the same answer shape: POST /api/laya runs laya.Router in-process, and POST /api/jev proxies to the TypeSafe System One endpoint with the key read from .env and never sent to the browser.

How do I deploy it on a server?

Run the container with docker run -d --name arena --restart unless-stopped -p 80:8732 --cpus 1 --memory 512m -e TYPESAFE_API_KEY=your_key_here ghcr.io/promptengineer48/laya-vs-jev-arena:latest, then open http://YOUR_SERVER_IP. DEPLOY.md covers the VPS, HTTPS, and key protection.

CommunityJev Demos & Experiments

Laya Vs Jev Arena

Side-by-side game harness where Laya running locally and Jev over API answer the same typed questions for snake, kombat, runner, flappy, tetris, and human-versus-model matches.

See highlights

Open github.com

Stars: 30
Language: JavaScript
License: MIT
Published by: Community
Author: PromptEngineer48
Last verified: Oct 2026

PromptEngineer48/laya-vs-jev-arena

User repository on GitHub

View on GitHub

Laya (open source, local) vs TypeSafe Jev (API): two AI models race in Snake and fight in a Mortal-Kombat-style arena. Every move is a real model decision.

Forks: 11
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

Two AI models go head to head in a snake race and a Mortal-Kombat-style fight. Every move is a real decision from the model — nothing is scripted. Laya — open source, Apache 2.0, runs locally on your machine · GitHub · Hugging Face. Jev — TypeSafe's closed model, runs over their API · announcement. Both take the same input and answer the same typed questions, so the only thing that changes between the two sides is the model. From the Prompt Engineer 48 YouTube video.

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

Tagscommunity jevbest catalog-import awesomejev

Works forJev for Classification

Used inJev for Software & DevTools

Languages: JavaScript

Highlights

Same state and questions to both models, with shared pace, warm-up, fixed rounds, and exported score and latency stats.
Snake race plus kombat ring with frame data, plus runner, flappy, and tetris on shared agent code.
Human keyboard options on every arena, including two-player matches on one keyboard.
Docker container and Python server.py proxy that keeps the Jev key server-side while Laya runs in-process.

Watch out

Jev latency includes your network path, the first local-model call loads checkpoints during priming, and game tabs must stay in the foreground or the browser throttles them. 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

live-jev

Browser top-down driving sim that sends sensor JSON to Jev every 200ms with four typed driving questions in one call, then applies confidence gating and reflex rules to steer.