Five labs, five minds: building a multi-model finance drama on small models
A hackathon project turns Thousand Token Wood into a game where one player manipulates an economy run by four different small models.
Intelligence analysis by GPT-5.4 Mini
This follow-up builds a playable finance drama on small models: the user becomes a shadow patron, while each creature is driven by a different lab’s model. The article focuses on the serving, security, and memory tricks needed to keep a heterogeneous agent economy stable.
A person built a tiny pretend town where little animal traders are each guided by a different brain. The player acts like a secret money boss, giving hints and loans, while the town’s rules and memory keep everyone from acting the same way.
Analysis
What changed in v2
The first Thousand Token Wood was a sandbox where five woodland creatures traded on one fine-tuned small model. Version 2 turns it into an interactive game: the user acts as the Patron of the Wood, lending, bribing, spreading tips, shorting the market, and forming alliances while a magistrate watches for abuse.
Heterogeneous models, one council
The main technical change is that each creature now runs on a different small model: gpt-oss-20b, MiniCPM3-4B, Nemotron-Mini-4B, and a fine-tuned Qwen 0.5B. The author argues that the point is not novelty, but realism: if the participants in a market behave differently, the market becomes more interesting. In practice, the hardest part was not model behavior but serving. The article says vLLM 0.22.1 needed nvcc at load time, so all models failed until the stack was moved to a CUDA devel image. Other issues were model-specific, like trust_remote_code for MiniCPM3.
A tolerant JSON parse-and-repair layer is central to the system. Because different models produce different formatting errors, the parser salvages what it can and keeps the simulation from breaking. That makes adding another model mostly a configuration task.
Secrets and memory
The game includes insider tips that may be true or false. A true tip can create profit, but only if the creatures do not see the hidden truth flag. The article treats this as a security problem, not a UI issue: the flag stays off-prompt, is stripped from public event records, and a test scans prompts every turn for banned tokens.
Relationships also persist. Creatures carry bounded sentiment toward the Patron and each other, and those feelings influence behavior. The system avoids prompt bloat by never dumping full history into the prompt. Instead, the model sees a short bucketed summary of the strongest feelings, while the full notes stay outside the prompt.
Results
The article reports a seeded run in which the council behaved as intended: the truth firewall leaked nothing, true tips created positive P&L, suspicious wins triggered the magistrate, and ruin from a margin call or default could banish a creature. The broader takeaway is that small models can work well when the system around them is structured carefully.
Key points
- The project turns Thousand Token Wood into an interactive finance game controlled by the player as a shadow patron.
- Each creature now uses a different small model from a different lab, making heterogeneity part of the design.
- The main engineering work was at the serving layer, including a CUDA/vLLM issue that blocked all models until fixed.
- Secret insider tips are kept off-prompt and verified with tests so the hidden truth cannot leak back out.
- Persistent relationships are summarized in bounded form to avoid prompt inflation and keep behavior testable.
If this approach holds up, small models could power richer multi-agent systems without needing giant models everywhere. The article suggests that strong structure, secret-handling, and bounded memory can make these systems feel alive and stay stable.
The setup still depends on careful serving and per-model fixes, so one bad deployment detail can break the whole council. The article also shows that secret leakage and prompt bloat are real risks unless they are enforced by tests and strict data-flow design.
