wiki/hackathon-context.md

metadata

title: Hackathon Context
tags:
  - context
  - hackathon
  - product
last_updated: 2026-06-16T00:00:00.000Z
confidence: high
sources:
  - https://huggingface.co/build-small-hackathon
  - >-
    official hackathon page (tracks, bonus quests, prize structure), re-fetched
    2026-06-05
  - >-
    https://build-small-hackathon-field-guide.hf.space/ (official Field Guide
    countdown, confirms the 23:59 UTC submission cutoff, verified 2026-06-15)
  - official Discord
  - HANDOFF.md
related_pages:
  - deployment-strategy.md
  - system-overview.md
  - model-selection-spike.md

Hackathon Context

Hatchimera — Voxel Pet Fusion is a submission for the Hugging Face Build Small Hackathon ("small models, big adventure"), track Thousand Token Wood. The hack window is open and the project is in active development.

Timeline

Official milestones are in UTC; the project is developed in Taiwan time (Asia/Taipei / UTC+8).

Milestone	UTC	Taiwan time (UTC+8)
Hack window starts	2026-06-05 Fri 00:01 UTC	2026-06-05 Fri 08:01
Submission deadline	2026-06-15 Mon 23:59 UTC	2026-06-16 Tue 07:59

Before the deadline, the Hugging Face Space, a short demo video, and a social media post must all be locked. For an authoritative cutoff, defer to the official page or dashboard. Winner announcement date: listed as TBD.

Tracks

The official page lists two tracks:

• Backyard AI: solve a real problem for someone you actually know. Scored on problem specificity, real user adoption, fit within constraints, and polish.
• Thousand Token Wood: build something delightful that wouldn't exist without AI. Scored on delight, load-bearing AI, concept originality, and polish.

Selected Track: Thousand Token Wood

This track calls for a fun, strange, shareable AI toy, tiny game, or art experiment. The key requirement: AI must be the thing doing the interesting work — not just assisting development, generating names, or writing a bio.

Scoring criteria:

• Delight: does the result make people want to show it to a friend?
• AI is load-bearing: is AI the core of the experience?
• Originality of concept: is the concept distinctive?
• Polish of the Gradio app: is it complete, responsive, and demo-ready?

Bonus Quest Targets (merit badges; 6 total)

Official framing (Field Guide): "No prize — just extra credit and bragging rights. Tag the ones your build earns." Descriptions are the official wording, verbatim from the Field Guide:

• Off the Grid (Local-first): "No cloud APIs. The whole thing runs on the model in front of you."
• Well-Tuned (Fine-tuned): "Your app uses a fine-tuned model you've published on Hugging Face."
• Off-Brand (Custom UI): "A custom frontend that pushes past the default Gradio look (hint: see gr.Server)."
• Llama Champion (llama.cpp): "Your model runs through the llama.cpp runtime."
• Sharing is Caring (Open trace): "You shared your agent trace on the Hub for everyone to learn from." Official clarification (Discord, 2026-06-08, organizer Dean + HF/Gradio's pngwn): the ideal is the trace from the model that runs inside your project at runtime, but any trace is accepted — local dev, coding-agent, or runtime — and a runtime agent trace is not required ("it's not necessary"). Share it on the Hugging Face Hub.
• Field Notes (Write-up): "You wrote a blog post or report about what you built and what you learned."

Submission Constraints

• Total model parameters <= 32B.
• App must use Gradio.
• App must be deployed as a Hugging Face Space.
• Submission must include a short demo video and a social media post.

Prize Structure (fetched 2026-06-05)

Total pool $48,000+, two tracks, ~29 awards. Per-track main prizes: 1st $4,000, 2nd $2,500, 3rd $1,500, 4th $1,000, Community Choice $2,000. Sponsor prizes ($22,000+): OpenBMB $10,000, OpenAI $10,000, NVIDIA (two RTX 5080), Modal $20,000 credits. Special awards include Bonus Quest Champion $2,000, Off-Brand $1,500, Tiny Titan (<=4B) $1,500, Best Demo $1,000, Best Agent $1,000, Judges' Wildcard $1,000.

Sponsor prizes (Field Guide, verified 2026-06-15)

Named sponsor awards with their official one-line requirements, verbatim from the Field Guide:

• Best MiniCPM Build — OpenBMB, $2,500: "Build with MiniCPM models."
• Best Use of Codex — OpenAI, $5,000: "Requires Codex-attributed commits in your connected GitHub repo or Space."
• Nemotron Hardware Prize — NVIDIA, RTX 5080: "Build with Nemotron models."
• Best Use of Modal — Modal, 10,000 credits: "Use Modal for the development or runtime of your app, and note it in your Space README."

These per-award amounts come from the current Field Guide; the sponsor-pool totals in the paragraph above were fetched 2026-06-05.

Project Fit

Hatchimera — Voxel Pet Fusion: draw a voxel pet from a text prompt → Splice two of them and one Gemma 4 12B call recombines their box geometry into a chimeric child (body from one parent, the signature feature from the other, plus a new mutation) → the child reveals on a 3D voxel stage and enters the family tree.

This fits Thousand Token Wood because AI is the genetics engine, not a garnish: both the drawn creature and the splice output cannot be reproduced from a menu, and they are the main thing on screen. The model emits raw box geometry (freeform), so it draws cross-form creatures a closed template vocabulary never could; the deterministic Tweak path (catalog parts, no model) keeps the cheap edits cheap. (See HANDOFF.md for the full rationale and the anti-regression note about the earlier "voice → pose enum" design that would have scored poorly on load-bearing AI.)