Five labs, five minds: building a multi-model finance drama on small models
Which summary reads better? Pick one — models revealed after.Both summaries are AI-generated.
Five labs collaborated to create a multi-model finance simulation where different small AI models interact as agents in an emergent economy, each representing distinct financial behaviors. Players act as shadow financiers, manipulating the market through tips, alliances, and trades while avoiding detection by a magistrate. The system relies on heterogeneous models from various labs, ensuring diverse decision-making and market dynamics.
Developers built version two of "Thousand Token Wood," an experimental finance game in which players act as a shadow financier—lending, shorting, bribing, and trading on insider tips while evading a pursuing magistrate—within an emergent woodland economy. The key engineering change runs each of the game's creature-agents on a different lab's small AI model, including OpenAI's gpt-oss-20b, OpenBMB's MiniCPM3-4B, NVIDIA's Nemotron-Mini-4B, and a fine-tuned Qwen 0.5B, so each behaves distinctly. The team found the main challenge lay at the serving layer rather than the modeling, solved through a tolerant JSON parse-and-repair system, while keeping insider-tip truth values hidden from the agents as a security requirement.