source: hugging face blog: thousand token wood: shipping a multi-agent economy on a 3b model
level: technical
thousand token wood is a simulation where five woodland creatures, each running on qwen2.5-3b, trade goods, gossip, and react to economic shocks. the model is served with vllm on modal, and a gradio app provides the interface. the project was built for the build small hackathon to explore what a 3-billion-parameter model can do in a multi-agent setting.
the initial version failed because production outran consumption, leaving no reason to trade. adding scarcity through diet variety, food spoilage, and a winter fuel crisis created the necessary pressure. the woodcutter's monopoly on firewood drove rising prices and a widening wealth gap. the model produced valid json on every call but made poor economic decisions until prompts were sharpened with role-specific instructions and worked examples.
a key feature lets players inject historical market events reskinned as woodland legends, such as a bank run causing an agent to dump assets and crash prices. prices were initially frozen because agents quoted reference values, so the system was changed to let market prices drift with supply and demand. over a fifteen-turn run, trades stayed active, prices moved realistically, and the gini coefficient rose from 0.14 to 0.38, with the woodcutter ending richest.
why it matters: it shows how small language models can power real-time multi-agent simulations when combined with careful system design, making complex economic modeling more accessible.
source: hugging face blog: thousand token wood: shipping a multi-agent economy on a 3b model