RawModels open-sources JEV Arena — a fighting game where decision models control fighters

🛠 RawModels open-sources JEV Arena — a fighting game where decision models control fighters

For each move, the model receives the battle state — health, distance, opponent's actions — and selects a technique. In the @arenaraw tournament, 18 models competed across 306 verified matches. The leader is d1-ultra from Liquid AI: Elo 1247, 94.1% win rate; Kev 4B lost all 34 matches.

🌍 The Bradley–Terry rating is built on replays with 95% confidence intervals, rather than vendor self-reports. The 'mental attacks' test showed that strength and resistance to prompt attacks are almost uncorrelated — Decider V1.2 (Perplexity) was defeated in 87.5% of positions, while Drex Max (Nace.AI) was not defeated in any.

👤 The arena can be opened in a browser to build your own fighter: model, character, and text-based strategy. Replays and built-in bots work without keys, while live matches require OpenRouter or RawModels API keys.

Source 1: https://www.rawmodels.ai/blog/jev-arena Source 2: https://github.com/rawmodels/jevarena