Multi-model finance drama uses four different small models in emergent economy game

The second version of Thousand Token Wood, a build-small hackathon project by Lester Leong, transforms a weather-god sandbox into an interactive finance game. You are the Patron of the Wood, a shadow financier who lends at interest, whispers tips (true or false), shorts markets, bribes, and brokers alliances, while a magistrate hunts you for illegal trading. The woodland creatures remember your actions and scheme back. The key innovation: each creature now runs on a different lab's small model — OpenAI's gpt-oss-20b, OpenBMB's MiniCPM3-4B, NVIDIA's Nemotron-Mini-4B, and a fine-tuned Qwen 0.5B. This heterogeneity creates a live argument rather than a scripted simulation, as the owl hoards differently than the fox speculates.

Standing up four distinct small models on one platform revealed that the friction lies almost entirely at the serving layer, not the modeling layer. All models used vLLM 0.22.1, which required a CUDA devel image to avoid a shared 'nvcc not found' error. Model-specific quirks were one-line config fixes. The real enabler was a tolerant JSON parse-and-repair layer that every model's output flows through, ensuring the simulation never crashes on malformed responses. To enforce information asymmetry, the truth flag for insider tips is kept entirely off-prompt and stripped from public event records; a test scans every creature's prompt for banned tokens. Memory is bounded by persistent sentiment scores nudged by events, avoiding prompt bloat. The result is a practical architecture for running heterogeneous, resource-constrained multi-agent systems.