Independent developer Lester Leong has created Thousand Token Wood Sim v2, an unscripted financial simulation game powered by five small language models spanning four distinct AI labs. By combining models under 32 billion parameters—including OpenAI's gpt-oss-20b, OpenBMB's MiniCPM3-4B, NVIDIA's Nemotron-Mini-4B, and a fine-tuned Qwen 0.5B—the project demonstrates how heterogeneous small models create emergent market behaviors. The entire multi-agent council operates on a single 24GB L4 GPU hosted on Modal, demonstrating cost-effective deployment strategies for complex AI simulations.
Overcoming Serving Layer Friction Across Heterogeneous Models
Rather than running a single model with varying system prompts, the project deliberately mixes models from different research labs to introduce true behavioral variety into the market dynamics. Because each lab's model exhibits unique fine-tuning biases and formatting quirks, the simulation achieves a genuine diversity of strategic behaviors among its trading agents.
Running four distinct model architectures on a single platform required overcoming technical hurdles at the serving layer. Using vLLM 0.22.1, the developer resolved loading failures by adopting a CUDA development base image containing the nvcc compiler required for JIT compilation. Additionally, a unified JSON parse-and-repair layer was introduced to automatically salvage malformed model outputs caused by differing tokenizer habits, preventing simulation crashes.
Key Architectural and Performance Results
- Fine-tuned 0.5B Qwen model achieved 100% valid offers and 0% self-buys, successfully outperforming its 3B teacher model in strict format adherence.
- Zero prompt leaks occurred across all automated test scans by keeping hidden state flags entirely off-prompt.
- Native MXFP4 quantization enabled OpenAI's 20B model to fit easily within a single 24GB L4 GPU footprint.
- Bounded memory summaries compressed historical interactions into simple integer-derived sentiment scores, preventing context window inflation.
Hardened Prompt Firewalls and Bounded Memory
To preserve game mechanics surrounding insider trading and rumors, the developer implemented strict data flow firewalls. Because small language models tend to repeat context provided in their prompts, secret information such as tip validity flags were isolated entirely from the prompt context and stored on an external ledger. Automated prompt-scanning test suites verified that private information never reached the agents or public event summaries.
To maintain long-term memory without drowning small model context windows, agent relationships are tracking mechanically using integer sentiment scores. These scores are converted into concise, bounded summary lines (such as feeling wary or warm toward another agent) rather than appending raw history. Hostile agents deterministically adjust loan rates or refuse trades, creating observable emergent behaviors without bloating the prompt.