A newer version of the Gradio SDK is available: 6.22.0
Build Small Hackathon — Requirements
Hugging Face × Gradio. "Build something small, local, and yours." Source: Field Guide · HF Org
🎯 Our target track: Thousand Token Wood (the whimsical track).
Note: "Thousand Token Wood" is the name of the whimsical track — there is no literal 1000-token budget. The only hard model constraint is ≤ 32B parameters per model (see REQ-01).
Key facts
| Final deadline | June 15, 2026 |
| Model cap | ≤ 32B parameters (per model) |
| Prize pool | $48k+ cash, plus 20k Modal credits, 2× NVIDIA RTX GPUs, ChatGPT Pro |
| Ways to win | 29 |
| Framework | Gradio app, hosted as a Hugging Face Space in the official Build Small org |
The big idea
The future of AI doesn't have to live in someone else's data center. Build Small is a return to small, local, tinkerable models — open weights you can read, run, and reshape, everything under 32B parameters, running on hardware you actually own. Less API bill, more workshop.
Tracks — two trails, equal prize pool
🌲 Thousand Token Wood — the Whimsical Track (OUR TRACK)
Whimsical, delightful, AI-native apps that push the boundaries of fun. Wander somewhere stranger and show off what small models can dream up.
Examples:
- Interactive AI games
- Out-of-the-box entertainment tools
- A desktop pet that lives on your machine
- A text-adventure dungeon master
What judges look for: genuine delight, the AI being load-bearing (not auxiliary), originality, and Gradio app quality.
🏡 Backyard AI — the Practical Track (for reference)
Practical, problem-solving apps built to improve daily life — for you or someone close to you. Useful things that run on hardware you own.
Examples: a custom storybook generator for a child, a personal study tutor, a receipt/bill parser, an on-device document assistant.
Entry criteria (REQ-01 → REQ-06)
Every submission needs all six.
- REQ-01 — Stay under 32B. Every model must be under 32B parameters. Combine several small models if you like — but each one's total parameter count must stay below the cap.
- REQ-02 — Ship a Gradio app. Deploy your project as a Gradio App inside the official Build Small org on Hugging Face. Docker is fine, as long as the interface is a Gradio Space.
- REQ-03 — Record a demo. Submit a demo video showing your app working — so judges can evaluate it even if GPU or API limits stop a live run.
- REQ-04 — Post it. Create one social-media post showcasing your app, and link to it from your Space README.
- REQ-05 — Mind the GPU limit. Submit as many apps as you like. If you rely on the provided Zero GPU resources, you're limited to 10 Zero GPU apps per user.
- REQ-06 — Tag your README. Add tags for the tracks and badges you want to be considered for to the YAML block at the top of your README, plus a short write-up of the idea and tech.
Prizes — 29 ways to win
General track prizes (awarded per track — so these apply to Thousand Token Wood)
| Place | Prize |
|---|---|
| 1st | $4,000 |
| 2nd | $2,500 |
| 3rd | $1,500 |
| 4th | $1,000 |
| Community Choice | $2,000 |
Sponsor prizes (each has its own criteria)
- OpenBMB — Best MiniCPM Build: 1st $2,500 / 2nd $1,500 / 3rd $1,000.
- Qualify: build with MiniCPM models.
- Pool is split per track ($5,000 to Backyard AI, $5,000 to Thousand Token Wood); placements awarded within each track.
- Vision (MiniCPM-V) and omni (MiniCPM-o) variants qualify, not just the text models.
- OpenAI — Best Use of Codex: 1st $5,000 / 2nd $3,000 / 3rd $1,000 ($10,000 total, across all submissions).
- Qualify: Codex-attributed commits in your connected GitHub repo or Space.
- Holistic use (fine-tuning, complex agents) ranks higher than light use.
- NVIDIA — Nemotron Hardware Prize: RTX 5080 for "Best space" + RTX 5080 for "Community engagement" (2 GPUs total, across all submissions).
- Qualify: build with Nemotron models. "Best space" judged by NVIDIA team; "community engagement" judged on likes & interactions.
- Modal — Best Use of Modal: 1st 10,000 / 2nd 7,000 / 3rd 3,000 credits ($20,000 in credits total).
- Qualify: use Modal for development or runtime, and note it in your README. Inference, fine-tuning, batch jobs, and sandboxes all count.
Bonus badges (tag them in your README)
- Off Brand — $1,500: best custom UI that pushes past the default Gradio look.
gr.Serveris your friend — go well beyond stock components. - Tiny Titan — $1,500: best app on a genuinely tiny model. Models must be ≤ 4B parameters; biggest impact from the smallest weights wins.
- Best Demo — $1,000: the full package — great app, demo video, and social post. Storytelling counts as much as the build.
- Best Agent — $1,000: the best agentic app. Multi-step tool use and planning, all under the 32B cap.
- Bonus Quest Champion — $2,000: the most bonus criteria met across the board. Ties go to the most ambitious, highest-quality submission.
- Judges' Wildcard — $1,000: for the entry that's amazing but fits no category. No entry needed — every submission is in the running.
Prizes stack. One Space can place in a track, win sponsor prizes, and collect bonus badges all at once.
How to submit
- Meet the criteria — double-check the entry rules (REQ-01→06) and any prize criteria you're targeting.
- Join the org — join the Build Small org on Hugging Face.
- Upload your Space — upload your submission as a Gradio Space inside the org.
- Record a demo — film a demo selling your Space (no humility). Put it on YouTube, upload to the Space, or host it publicly.
- Post on social — share one post about your build.
- Update your README — add links to the post + demo video, tags for tracks + badges in the YAML block at the top, and a short write-up of the idea and tech.
Submit at: https://build-small-hackathon-field-guide.hf.space/submit
FAQ highlights
- What does "under 32B" mean? Every model your project depends on must have under 32B total parameters (not just active params). You can combine several models (e.g. a 14B text + 7B speech + 12B image model) as long as each stays under the cap.
- Must I use a sponsor's model? No. General tracks accept any model under 32B, including from non-sponsoring orgs. Sponsor prizes require their own model/platform per their criteria.
- Hosted API vs local? Hosted APIs are allowed (some partners offer free hosted access; HF inference providers offer more). Some prizes require local execution to be eligible.
- GPU limit? If using the provided Zero GPU resources: max 10 Zero GPU apps per user. Otherwise run on consumer hardware or Modal credits.
- Multiple apps? Yes — submit as many as you like; each is considered.
Partners / kit (small-AI models worth reaching for)
- OpenBMB — MiniCPM family: tiny text · vision · audio · omni models (1B–8B). E.g. MiniCPM-V 4.6 (~1.3B, strong OCR/doc understanding).
- Black Forest Labs — FLUX.2 Klein: text-to-image & precise image editing at 4B / 9B.
- OpenAI · Codex — Codex coding agent (GPT-5.5) with GitHub, Figma & Hugging Face plugins.
- NVIDIA — Nemotron 3 family: Nano · Omni · ASR · Parse · Embed (Nemotron Parse: sub-1B structured doc extraction).
- Modal — serverless compute for inference, training, batch & sandboxes.
- JetBrains — Mellum 2: 12B MoE coding models (Thinking & Instruct).
- Cohere Labs — Cohere Transcribe (ASR) and Tiny Aya multilingual models.
© 2026 Build Small · ≤ 32B params · open weights · run it yourself