A multi-agent market runs 4 rival small models — friction was just a missing CUDA toolkit
hugging face's 'thousand token wood' simulation runs heterogeneous agents on gpt-oss-20b, minicpm3-4b, nemotron-mini-4b, and qwen. the experiment proves that distinct model architectures create emergent economic behaviors, while vllm configuration remains the primary engineering bottleneck.
hugging face released a field report on 'thousand token wood v2', a multi-agent economy where participants operate on distinct small models: gpt-oss-20b (mxfp4), minicpm3-4b, nemotron-mini-4b, and a fine-tuned qwen 0.5b. the project utilized these divergent models—trained on different data with different post-training—to simulate a genuine market where agents 'argue' via trading mechanics rather than following a script. engineering friction was minimal despite the model variety; the major blocker was a vllm dependency issue where the cuda toolkit (nvcc) was missing from lean base images, requiring a switch to a cuda devel image. the project also implemented strict information security protocols, ensuring that secret insider tips provided to agents are stripped from prompts before processing to prevent data leakage, highlighting the architectural requirements for secure multi-agent environments.