Five labs run multi-model finance game on small models

Five labs' small models power a finance game where five agents trade in a forest economy. Each agent uses a different model: GPT-oss-20b, MiniCPM3-4B, Nemotron-Mini-4B, and a fine-tuned Qwen 0.5B. Players act as financiers, lending, shorting, and influencing agents while agents remember past interactions and scheme back. The system runs on tolerant JSON parsing, showing friction is at serving layer, not modeling. The setup enables diverse agent behaviors and emergent market dynamics.