Build Small Hackathon 2026

Build Small was a Hugging Face × Gradio hackathon, held June 5–15, 2026, with a premise I liked immediately: build a real, tinkerable app powered entirely by open-weight models under 32B parameters. No frontier API doing the actual work — small, local models carrying the whole load, with a $48K+ prize pool and a sponsor lineup spanning OpenBMB, Black Forest Labs, OpenAI, NVIDIA, Modal, JetBrains, and Cohere Labs.

My entry was CXR Draft Auditor, a research and educational quality-assurance tool that gives a chest X-ray draft impression a transparent second read. Two tiny 4B models — a twice-fine-tuned MedGemma that grounds the image into labeled findings with bounding boxes and an NVIDIA Nemotron-3 Nano 4B that parses the draft into the same labels — feed a deterministic, model-free comparator that does the only judging, and both run inside the live Hugging Face Space, comfortably within the hackathon's parameter budget.

Small models are usually the compromise; Build Small treated them as the point. Working under a hard sub-32B budget is exactly what pushed the design I ended up most proud of — two narrow models, each doing the one job it is good at, feeding a judgment layer with no model in it at all. The full build story, from the no-PhysioNet data work to the QLoRA fine-tuning to the evaluation-integrity bug that nearly shipped the wrong model, is in my Field Notes write-up.

Below is the certificate of participation from the hackathon.