Five labs, five minds: building a multi-model finance drama on small models

Source: HuggingFace Date: 2026-06-06 URL: https://huggingface.co/blog/build-small-hackathon/thousand-token-wood-sim-v2

Summary

A HuggingFace Build-Small hackathon entry that simulates an emergent financial market using four heterogeneous small models — GPT-OSS-20B, MiniCPM3-4B, Nemotron-Mini-4B, and a fine-tuned Qwen 0.5B — each playing distinct agent roles. The project’s headline result is that a 0.5B model fine-tuned for structured output achieved 100% trade-validity, closing a reasoning gap that naive prompting could not close at that scale. The authors found that friction in multi-model systems lives almost entirely at the serving layer, not the modeling layer, and that information firewalls between agents need test coverage, not prompt-based restrictions.

Implications

• Small-model economy. The Qwen 0.5B result is the sharpest data point here: targeted fine-tuning on a structured task can rival larger models on that task, reinforcing the “right model for the job” routing logic that enterprise builders are starting to operationalize.
• Agentic-engineering patterns. The firewall-as-security-property framing is worth tracking — the team is describing inter-agent information isolation as an engineering concern with test coverage requirements, not a prompt-craft concern. That’s a maturity signal.
• Multi-model orchestration. Using genuinely different models to create behavioral diversity (rather than varied system prompts on one model) is an emerging pattern with real cost and complexity implications; this project is a small-scale proof-of-concept for heterogeneous agent fleets.