The two hard constraints, acceptance criteria, timeline, and judging.
Everything else is advice. These two are non-negotiable.
Every AI model you use in your application must have 500 million parameters or fewer. This is measured against the model's architecture/parameter count, not its quantized file size. You may use multiple models in one project, as long as each individual model is under the cap.
Tip: a 22M MiniLM often does more than a 400M model you can't fit. Shrink the problem, not just the model.
Models must run locally within your app. You may not call OpenAI, Anthropic, Google, the Hugging Face Inference API, Ollama cloud, or any other remote inference-as-a-service. All inference happens on the participant's own machine/device at runtime.
You may use non-AI APIs: maps, auth, payments, storage, public data sources, etc.
Your project is accepted for judging if and only if all of the following are true:
This project was submitted to the ryze.ai hackathon by [NAME OF SUBMITTER].
(Replace [NAME OF SUBMITTER] with the submitter's name.)
As long as these are met, your project is accepted. The 500M / no-API rules are then enforced during judging — projects found to violate them are disqualified and forfeit the prize.
All times UTC.
| Date | Milestone |
|---|---|
| Sep 8, 2026 | Registration opens |
| Sep 28, 2026 | Registration closes |
| Sep 29, 2026 | Kickoff livestream, hacking begins |
| Oct 27, 2026 | Submissions due |
| Nov 12, 2026 | Winners announced |
Submissions are scored and ranked — the top entry wins the $1,000.
| Criterion | Weight | Looks for |
|---|---|---|
| Impact & Utility | 25% | Solves a real problem, genuinely useful |
| Creativity & Novelty | 25% | Original approach, surprising small-model win |
| Technical Execution | 30% | Solid integration, clean code, on-device inference |
| Presentation & Demo | 20% | Clear, engaging, polished |