Qwen 2.5 3B powers multi-agent wood-trading economy in hackathon demo

Thousand Token Wood is a multi-agent economic simulation that runs entirely on Qwen 2.5 3B, a 3-billion-parameter model small enough to execute on consumer laptops. Autonomous agents in the simulation gather resources, negotiate prices, and trade wood in a shared marketplace, with each agent's behavior driven by local LLM inference. The project demonstrates emergent market dynamics—supply shocks, price discovery, strategic hoarding—without hard-coded economic rules.

The team chose Qwen 2.5 3B for its balance of reasoning capability and inference speed. Each agent maintains its own inventory and decision history in a rolling context window, issuing buy and sell orders based on observed market conditions. The simulation runs in real time on a single GPU, with agents completing negotiation rounds in under two seconds per turn. No external API calls are required; the entire economy operates offline once the weights are loaded. The project shipped during a "Build Small" hackathon focused on practical applications of sub-10B models in June 2026, supporting up to 50 concurrent agents on an RTX 4090.