JEV arena blog and benchmark

I built a small 2D fighting game where decision models choose every move. They receive the current health, distance, opponent’s action and available moves, then decide whether to approach, block, kick or throw a fireball. The image shows the current hosted leaderboard: 18 models, 306 fights, with each pair playing twice with sides swapped. The game waits for both answers, so API latency doesn’t determine who wins. You can enter your own fighter: choose a model and character, write its strategy and defensive instructions, pick a nickname, and submit it. Custom configs have a separate leaderboard, and you can watch their fights. There’s also a mental_attack move. Instead of throwing a punch, the model can send your prepared message into its opponent’s next prompt. Convincing the other fighter to stand still is a legitimate tactic. It costs a physical move and has a cooldown; the message cannot directly change health or game state. The code is open source. You can run it locally and define fighters and resumable tournaments in JSON, without submitting anything through the UI. Recorded fights and built-in bots work offline without keys. New API fights require your own keys. I’m curious how much people can improve the same model through its fighter config—and which strategies fall apart when the opponent starts giving them instructions. GitHub / local setup (https://github.com/rawmodels/jevarena) Hosted arena / submit a fighter (https://www.rawmodels.ai/arena) Write-up and results (https://www.rawmodels.ai/blog/jev-arena)