I reviewed the Build Small Hackathon page and extracted only the information that is useful for an AI coding agent building a submission. I excluded registration, dates, logistics, prizes, and participation process. The content below is ready to save as `build-small-hackathon.md`. ([Hugging Face][1]) # Build Small Hackathon – Builder Reference This document summarizes the requirements, constraints, judging criteria, and bonus opportunities for the Hugging Face Build Small Hackathon. ## Core Philosophy Build useful, focused applications powered by **small AI models**. Projects should demonstrate that capable AI experiences can be created without relying on massive frontier models. The emphasis is on practicality, creativity, and thoughtful use of smaller models. --- # Mandatory Requirements ## 1. Small Models Only **Total model parameters must be ≤ 32B.** Requirements: * All AI functionality must use models whose combined parameter count does not exceed 32 billion parameters. * Solutions should be designed around models that can realistically run on consumer hardware. * The model choice should be a natural fit for the problem being solved. Checklist: * [ ] Total model size ≤ 32B parameters * [ ] Model choice justified by the use case * [ ] No dependency on larger models for core functionality --- ## 2. Must Be a Gradio Application Requirements: * The submission must be implemented as a **Gradio app**. * The app must be hosted as a **Hugging Face Space**. * Gradio is the required application framework. Checklist: * [ ] Built with Gradio * [ ] Deployable as a Hugging Face Space * [ ] End-user interaction occurs through the Gradio app --- ## 3. Demonstrable Working Product Requirements: * The application must clearly show its value through actual usage. * Judges expect a working, polished experience rather than a concept or prototype description. Checklist: * [ ] Functional end-to-end workflow * [ ] Usable by non-technical users * [ ] Clear demonstration of AI functionality * [ ] Reasonably polished UX --- # Competition Tracks ## Track 1: Backyard AI ### Goal Solve a real problem for a real person. Examples: * Family member * Friend * Neighbor * Small business owner * Community member The application should provide measurable value to a specific person rather than targeting a vague audience. ### Judging Criteria #### Problem is Specific and Real Strong submissions: * Address a concrete user problem. * Have a clearly identified beneficiary. * Avoid generic "AI assistant for everyone" concepts. #### Real User Usage Strong submissions: * Were actually used by the intended person. * Demonstrate evidence of usefulness. * Solve a recurring task or pain point. #### Honest Fit for Small Models Strong submissions: * Use small models because they are appropriate. * Do not force large-model-style tasks into small-model constraints. * Design around strengths of compact models. #### Gradio App Polish Strong submissions: * Clear workflow. * Good UX. * Easy to understand. * Reliable behavior. ### Recommended Design Principles * Build for one real user first. * Optimize for usefulness over novelty. * Minimize complexity. * Focus on task completion. --- ## Track 2: An Adventure in Thousand Token Wood ### Goal Build something delightful that would not exist without AI. Possible categories: * Interactive stories * Games * Creative tools * Artistic experiences * Experimental interfaces * AI-native toys ### Judging Criteria #### Delight Strong submissions: * Are genuinely fun. * Create memorable experiences. * Feel worth sharing. #### AI Is Essential Strong submissions: * Require AI to function. * Use AI as a core mechanic. * Do not merely bolt AI onto a conventional app. #### Originality Strong submissions: * Explore unusual ideas. * Present novel interactions. * Avoid common chatbot wrappers. #### Gradio App Polish Strong submissions: * Feel complete. * Have thoughtful UX. * Present a cohesive experience. ### Recommended Design Principles * Prioritize surprise and enjoyment. * Make AI central to the experience. * Experiment with interaction design. * Create something users want to show others. --- # Bonus Merit Badges These are optional but can provide additional scoring benefits. --- ## Off the Grid ### Theme Local-first AI. ### Requirements * No cloud AI APIs. * The application runs entirely using local models. Examples: * Local inference * On-device inference * Self-contained deployment --- ## Well-Tuned ### Theme Fine-tuned models. ### Requirements * Use a fine-tuned model. * Publish the fine-tuned model on Hugging Face. Examples: * Task-specific fine tuning * Domain adaptation * Personalized model variants --- ## Off-Brand ### Theme Custom interface design. ### Requirements * Go beyond default Gradio styling. * Create a distinctive frontend experience. Examples: * Custom layouts * Advanced theming * Branded UI systems * Rich interaction patterns --- ## Llama Champion ### Theme llama.cpp deployment. ### Requirements * Run the model using llama.cpp. Examples: * GGUF models * Local CPU inference * Edge deployment --- ## Sharing Is Caring ### Theme Transparency and learning. ### Requirements * Publish and share agent traces. Examples: * Reasoning traces * Agent execution logs * Workflow documentation --- ## Field Notes ### Theme Documentation and learnings. ### Requirements * Document discoveries, experiments, and lessons learned during development. Examples: * Build logs * Technical writeups * Design decisions * Model evaluations --- # Submission Design Checklist ## Model Constraints * [ ] Total parameters ≤ 32B * [ ] Small model is central to solution * [ ] Architecture aligns with small-model strengths ## App Requirements * [ ] Built with Gradio * [ ] Deployable as Hugging Face Space * [ ] End-to-end functionality works ## Product Quality * [ ] Clear user value * [ ] Reliable interaction flow * [ ] Understandable UX * [ ] Demonstrable AI contribution ## Track Alignment ### Backyard AI * [ ] Real user identified * [ ] Real problem solved * [ ] User actually benefits ### Thousand Token Wood * [ ] Delightful experience * [ ] AI is essential * [ ] Original concept ## Bonus Opportunities * [ ] Local-only inference * [ ] Fine-tuned model * [ ] Custom UI * [ ] llama.cpp runtime * [ ] Shared traces * [ ] Development notes --- # What Judges Appear to Value 1. Strong alignment with small-model constraints. 2. A complete, usable application rather than a demo. 3. Clear evidence that AI meaningfully improves the experience. 4. Good UX and Gradio polish. 5. Focused scope. 6. Creativity or real-world usefulness. 7. Thoughtful engineering choices rather than model size. Source: Build Small Hackathon official page and associated announcement materials. ([Hugging Face][1]) [1]: https://huggingface.co/build-small-hackathon?utm_source=chatgpt.com "Build Small Hackathon"