Model fight club

What it does

Model Fight Club puts 4 open-weights LLMs (Kimi, DeepSeek, Qwen, GPT-OSS) in a real-time arena on Mozaik. Round 1: all four answer the same prompt at the same instant, fastest takes the crown. Round 2: the winner's answer becomes the new prompt and every fighter rebuts or defends it — concurrently. Every run streams a live, replayable trace to Mozaik Cloud and costs ~$0.0007.

How the agents run concurrently

One message.sent event fans out to 4 fighter agents in a shared AgenticEnvironment; each handler awaits only its own streaming LLM call, so all four infer at once while per-token fighter.delta events interleave on the bus. They share FightState (prompt, results[], rebuttals[], round2 flag), and coordination is causal, not scripted: the 4th model.answer fires round2.start carrying the champ's answer, which all fighters react to in a second concurrent race. A Scoreboard observer ranks both rounds; a cloud-mirror participant streams everything live. A fighter that errors still posts model.answer with finishReason: "error", so one lab's outage never kills the bout.