Laya vs Jev arena
Browser arena where local Laya and hosted Jev race in Snake and fight in a Mortal-Kombat-style game, with every move a real decision.
Overview
Laya vs Jev Arena pits two typed decision models against each other: Laya, open source and running locally, and Jev, TypeSafe's closed model over its API. They race in a snake game and fight in a Mortal-Kombat-style arena, and every move is a real model decision rather than a script. Both sides receive the same state and typed questions, so the only difference is the model. It accompanies a Prompt Engineer 48 YouTube video.
A Python server serves the pages and exposes two backends with the same request and response shapes: Laya in-process via its router, and Jev proxied so the key never reaches the browser. You can choose a model per side, including a human player, change the seed for a different apple layout shared by both sides, and export JSON with score, decisions per second, p50 and p95 latency and tokens. The first Laya call takes 30 to 60 seconds to load checkpoints, done before the clock starts. A Docker image lets you host a human-versus-Jev game. JevK5's README uses this arena for its Snake demo.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which direction should the snake turn next?(inferred) | — |
| Choice | Which fighting move should the character make?(inferred) | — |
How it works
- Choose a model for each side (Laya, Jev or human)
- Game sends the same state and questions to both backends
- Each model's choice drives its character
- Stats track score, decision rate and latency
- Export results as JSON
Strengths
- Same inputs for both models make the comparison fair
- Key stays server-side
- Seeded runs and exportable metrics
- Docker deployment
Caveats
- Game performance is not a general accuracy benchmark
- Browser tab must stay in the foreground
- Jev side needs a TypeSafe key
Getting started
git clone https://github.com/PromptEngineer48/laya-vs-jev-arena.git, pip install -r requirements.txt, add TYPESAFE_API_KEY to .env, python server.py 8740.At a glance
- Type
- Service
- Status
- prototype
- Language
- Python
- Runtime
- Python server and browser games
- Has code
- yes
- License
- MIT
Use cases and integrations
How curated lists describe it
Model comparison: races an open local model against Jev through Snake and a Mortal-Kombat-style arena, the same game code driving both.
yibie/awesome-jevLaya (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.
logicrw/awesome-jev-projectsLaya (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.
hellogumbo/awesome-jevA local open model and the hosted one play a snake race and a fighting game against each other, every move a real decision rather than a script.
MrJev/awesome-jev
Frequently asked questions
Is the comparison fair?
Both models get the same state, questions and seed; only the model differs.
Do I need a TypeSafe key?
Only for the Jev side; Laya runs locally with no key.
Why is the first Laya move slow?
It downloads and loads checkpoints once, which takes 30-60 seconds before play starts.
Related projects
Chrome T-Rex game where local Laya on MLX and hosted Jev choose moves side by side, with latency charts and crash replays.
ServiceprototypePythonBrowser Pac-Man race comparing hosted Jev and self-hosted Laya choosing routes from the same board state, after code removes fatal routes.
ServiceprototypeJavaScriptFifteen live arcade games where Jev and LLMs play each other, led by WikiRace in which Jev picks only real links.
ProjectbetaTypeScriptBrowser extension that lets Jev, local Laya or an OpenAI-compatible model take over a running ra2web (WannaFire) RTS match.
ProjectbetaTypeScriptOne-page web app where Jev plays chess against OpenRouter LLMs, Stockfish or you, showing the moves it weighed and their probabilities.
AppbetaJavaScriptPac-Man where Jev or other System One models choose each move at junctions, with live odds, fallbacks and a leaderboard.
ProjectbetaTypeScript