3Bモデルで経済を回せる?Thousand Token Woodは「イエス」と答える

thousand token wood uses qwen2.5-3b to power a real-time economy of five trading woodland creatures on modal, demonstrating that small models can sustain complex multi-agent systems. while the 3b model emitted valid json on 100% of calls, it initially struggled with economic judgment until scarcity mechanisms like spoilage and fuel crises were introduced. this highlights that with proper prompt engineering and system design, small models can handle agent logic effectively, though they may require constraints to avoid economic paradoxes. the project successfully recreated market events like bank runs and bubbles, signaling a shift towards using smaller, cheaper models for agent simulations rather than relying on expensive frontier apis.