Hugging Face × Gradio N 45.21 · W 122.6

BUILD SMALL

Build something small, local, and yours.

Registration’s closed and the jam is underway — this is your field guide. Everything you need to enter cleanly: the rules, the 29 ways to win, and the right kit for what you’re building.

June 15, 2026final deadline
≤ 32Bparams, max
$48k+prize pool
29ways to win
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, humming away on hardware you actually own. Less API bill, more workshop.

02 Pick a trail

Two tracks, one campsite

Solve a real problem, or wander somewhere weird. Both are equally celebrated — and carry the same prize pool.

THE PRACTICAL TRACK

Backyard AI

Practical, problem-solving apps built to improve daily life — for you or someone close to you. Useful things that run on hardware you own.

  • A custom storybook generator for a child
  • A personal study tutor
  • A receipt or bill parser
  • An on-device document assistant
THE WHIMSICAL TRACK

Thousand Token Wood

Whimsical, delightful, AI-native apps that push the boundaries of fun. Wander somewhere stranger and show off what small models can dream up.

  • Interactive AI games
  • Out-of-the-box entertainment tools
  • A desktop pet that lives on your machine
  • A text-adventure dungeon master
03 Trail rules

Entry criteria at a glance

The things every submission needs. Tick them off and you’re on the board.

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 frontmatter tags for the tracks and badges you want to be considered for, plus a short write-up of the idea and tech.

04 The prize table

29 ways to win

A $48k cash pool plus 20k Modal credits, two NVIDIA RTX GPUs and ChatGPT Pro — across track placements, sponsor challenges, and collectable bonus badges.

$48k
cash pool
+29
ways to win
GENERAL TRACK PRIZES · AWARDED PER TRACK
Backyard AI
Thousand Token Wood
SPONSOR PRIZES · OWN CRITERIA PER PRIZE
BONUS BADGES · TAP FOR DETAILS
$1,500
Off Brand

The best custom UI that pushes past the default Gradio look.

gr.Server is your friend — go well beyond the stock components and make it yours.

$1,500
Tiny Titan

The best app built on a genuinely tiny model.

Models must be ≤ 4B parameters. Biggest impact from the smallest weights wins.

$1,000
Best Demo

The full package: great app, great demo video, great social post.

Storytelling counts as much as the build — sell it across all three.

$1,000
Best Agent

The best agentic app.

Multi-step tool use and planning — all under the 32B cap, of course.

$2,000
Bonus Quest Champion

The most bonus criteria met across the board.

Ties go to the most ambitious, highest-quality submission despite the extra constraints.

$1,000
Judges’ Wildcard

For the entry that’s amazing but fits no category.

No entry needed — every submission is in the running. We know you’re out there.

05 Choose your kit

What are you building?

Tell us the shape of your idea and we’ll point you at the partners and models worth reaching for. Then dig into their pages for the full guide.

For a image / ocr app, reach for:

Read documents, understand photos, or generate & edit images.

MiniCPM-V 4.6 via OpenBMB

Vision-language at ~1.3B — strong OCR & document understanding.

Open
FLUX.2 Klein via Black Forest Labs

Generate and edit images locally at 4B / 9B.

Open
Nemotron Parse via NVIDIA

Sub-1B structured extraction from complex documents.

Open
06 The outfitters

Seven partners stocked the shed

Models, tools and compute from across the small-AI world. Tap any one for its full kit and support channels.

MiniCPM family — tiny, capable text · vision · audio · omni models (1B–8B).

1B–8B models

FLUX.2 Klein — text-to-image & precise image editing at 4B / 9B.

image gen

Codex coding agent (GPT-5.5) with GitHub, Figma & Hugging Face plugins.

coding agent

Nemotron 3 family — Nano · Omni · ASR · Parse · Embed.

model suite

Serverless compute for inference, training, batch & sandboxes.

compute

Mellum 2 — 12B MoE coding models, Thinking & Instruct.

12B MoE

Cohere Transcribe (ASR) and Tiny Aya multilingual models.

ASR · multilingual
FIND THE RIGHT ONE Use the kit recommender →
07 The trail map

How to submit

The markers between you and the finish line.

1

Meet the criteria

Double-check your build satisfies the entry rules and any prize criteria you’re targeting.

2

Join the org

Join the Build Small hackathon organisation on Hugging Face — your home base for the jam.

3

Upload your Space

Upload your submission as a Gradio Space inside the org.

4

Record a demo

Film a demo selling your Space — no humility. Put it on YouTube, upload it to the Space, or host it publicly.

5

Post on social

Share one post about your build on social media.

6

Update your README

Add links to the post and demo video, frontmatter tags for tracks + badges, and a short write-up of the idea and tech.

08 Field notes

Frequently asked

Every model your project depends on must have under 32B total parameters (not just active parameters). You can freely combine several models — say a 14B text model, a 7B speech model, and a 12B image model — as long as each one individually stays under the cap.
No. For the general tracks you can use any model you like, including ones from non-sponsoring orgs, as long as it stays under the 32B cap. Sponsor prizes each require their own model or platform and have their own criteria, listed on every prize.
No — you can use other providers’ models too. But the sponsor’s models must form a core part of the experience, since each sponsor is looking for the most interesting use of their own models.
Yes. The credits given to the first 1,000 registrants aren’t tied to the prize. As long as you use Codex during development and have Codex-attributed commits in the Space’s history or a linked GitHub repo, you’re eligible.
If you rely on the provided Zero GPU resources, you’re limited to 10 Zero GPU apps per user. Otherwise, run on consumer hardware or use Modal credits.
Yes. Several partners (like OpenBMB) provide free hosted API access, and Hugging Face inference providers offer plenty more options. That said, not every model is available via a cloud API, and some prizes require running locally to be eligible.
Yes — awards stack. One Space could place in a general track (say Backyard AI), win sponsor prizes (like OpenBMB’s Best MiniCPM Build, the OpenAI Codex prize, and Best Use of Modal), and still pick up bonus badges such as Tiny Titan if every model is ≤ 4B — all at once.
Yes — submit as many apps as you like, and each one will be considered.
Ship a Gradio Space in the org, record a demo video, post on social, and update your README with links and frontmatter tags. See the trail map above for the full sequence.