Updated October 21, 2026. The Nebius x NVIDIA Global AI Hackathon challenges developers to deliver a working AI application built using Nebius infrastructure and at least one NVIDIA open-source model. Unlike a pitch-only competition, the official rules require reproducible software, a public open-source repository and a short demonstration video. The submission deadline is October 30, 2026, at 10:00 a.m. Pacific Time.

This in-depth guide is grounded in the event's official rules. We separate requirements that can disqualify a submission from practical recommendations for creating a convincing project. Always check the live official rules for amendments; the organisers expressly reserve the right to change them.

1. Dates, organisers and the basic facts

MilestoneOfficial schedule (Pacific Time)
Submission opensAugust 26, 2026, 9:00 a.m.
Submission closesOctober 30, 2026, 10:00 a.m.
Judging beginsDecember 1, 2026, 9:00 a.m.
Judging endsDecember 15, 2026, noon
Winners announcedOn or around January 11, 2027, noon

The sponsor is Nebius B.V., based in the Netherlands; the administrator is Devpost, Inc. The rules state that no purchase or payment is necessary to enter or win, and paying does not improve winning chances. Registration takes place on Devpost through the Join Hackathon action, followed by a project submission during the stated window. Joining and submitting are distinct steps.

The official overview shows a substantial participant community, but a participant count must not be confused with the number of completed, eligible projects or with an individual entrant's probability of winning.

2. Eligibility: individuals, teams and organisations

The rules accept individuals who have reached the age of majority where they live, teams of eligible individuals and legally established organisations. Teams and organisations must appoint an authorised representative who submits and acts on their behalf. An eligible individual can belong to more than one team or organisation and may also enter individually, subject to the requirements and originality of each submission.

Eligibility is not universal. The official rules exclude certain places where US or local law prohibits participation or receipt of prizes and expressly mention Brazil, Quebec, Russia, Crimea, Cuba, Iran and North Korea, among other restrictions. They also exclude judges, organiser and promotional entities, relevant staff and relatives, and people with certain conflicts of interest. Review the precise language rather than assuming that an online event is available in every jurisdiction.

Make your representative and repository ownership clear before the final week. For a multi-person project, confirm who is authorised to submit, how prize distributions would work and who has the right to publish any existing company code.

3. Non-negotiable project requirements

Every entry must be a working software application running on Nebius Token Factory or Nebius AI Cloud, using at least one NVIDIA open-source model and fitting one of the four tracks. Under the rules, running on the platform means making a runtime call to the Token Factory inference API or deploying/running the solution using Nebius AI Cloud compute such as Serverless Jobs, Serverless Endpoints or DevPods.

Merely displaying Nebius or NVIDIA in a slideshow, mentioning them in a README, or using an unrelated model hosted elsewhere would not establish the required runtime integration. Your application needs to operate as described, install and run consistently on its intended platform and be available for evaluation.

You may submit a new project, or a pre-existing one if it was significantly updated after August 26, 2026. Explain the changes specifically in the submission; renaming an existing product or changing its homepage is not convincing proof of a substantive update. Third-party SDKs, APIs and datasets are allowed only where you have the necessary rights and comply with their terms.

4. Four tracks: choose the best fit

Coding and Agentic Engineering

Build coding agents or developer tools that write, run and test code using Token Factory. A useful design could take a small repository, identify a reproducible test failure, propose a patch, run a constrained test suite and show what changed. Engineering quality matters: isolate execution environments, limit shell permissions and produce an auditable record of actions rather than presenting unreviewed code execution as safe.

Best Apps and Agents

Build a practical application or agent that people would actually use. The rules specifically describe using Nemotron models on Nebius via Token Factory, suggesting larger reasoning models for difficult tasks and smaller variants for low-latency steps. Nebius Serverless Endpoints and Serverless Jobs are encouraged for appropriate workflows, but the track text says these two deployment options are not mandatory.

A clear entry has a real user, measurable outcome, useful interface and reliable error handling. A workflow that automates a narrow administrative task can be more persuasive than a generic assistant that tries to do everything.

Personal AI

The emphasis is on an always-on, private personal assistant with controlled data, persistent memory, reusable skills and chosen tool access. The official description refers to technologies including NVIDIA NemoClaw, OpenShell, Hermes Agent and Nebius Serverless. Treat privacy as an architectural claim that needs evidence: show where data is stored, who can read it, how permissions are granted and how the assistant's access can be revoked.

Physical AI

This category covers embodied and edge systems, robotics, IoT and agents working with the physical world. The track description mentions Nemotron, GROOT, Cosmos and Sonic, as well as simulation, sensor processing and asynchronous workloads on Nebius. There is a specific video rule: include at least one minute of real hardware or robot operation; if the project has no hardware component, show key application modules operating for that portion instead.

5. The required Devpost submission package

This is one of the most important parts of the event. According to the rules, the submission must contain:

  1. A working project satisfying the Nebius and NVIDIA technology requirements, submitted to the chosen track.
  2. A URL for a working demo, hosted application or test build, except that Physical AI submissions follow their special hardware access provision.
  3. A plain-language explanation of features and functionality.
  4. A public source-code repository on GitHub, GitLab or Bitbucket with code, assets and instructions needed for the project to function.
  5. A visible open-source licence file, such as MIT, Apache 2.0 or MPL 2.0, and a README explaining setup and execution.
  6. An explanation of which NVIDIA model is used, how Nebius Token Factory or AI Cloud is involved and what other relevant Nebius components do.
  7. A publicly accessible YouTube demo video shorter than three minutes, showing the project working on its intended device.
  8. Track selection, feedback on the relevant tools and, for old projects, an explanation of the substantial changes made during the event window.

The project must be made available to judges for free and without restrictions until judging ends. If a demo is private, the testing instructions need credentials that genuinely work. The rules allow judges to assess submissions using descriptions, images and videos without testing the app in every case, so the supporting materials matter even when the software is solid.

Submissions must be in English, or have English translations of the demonstration, written description, testing instructions and other submitted material. Music, third-party trademarks and copyrighted footage in videos require authorisation from rights holders.

6. Make the public repository genuinely reproducible

Because a public, open-source repository is mandatory, check repository policy early, particularly if you are building on a company's internal platform. It is not enough to paste a licence name into the article about your project while leaving the code closed.

As an editorial recommendation, structure your README around prerequisites, supported versions, architecture, environment-variable names (without secrets), installation commands, how to obtain model access, one working example and how to run tests. Provide sample data that can legally be published. Pin critical dependency versions and include a short explanation of model input/output contracts.

For AI agents, show what the system can do when API calls fail, what happens with malformed model responses, whether external actions require approval and how long-running tasks can be cancelled. These are engineering recommendations, not extra eligibility clauses.

7. Judging: two stages and four equal criteria

Stage one is pass/fail. The submission must be viable, genuinely fit its track and meaningfully use the required APIs, SDKs and models; superficial rebranding of an unrelated project is explicitly insufficient.

Entries passing this filter are evaluated on four equally weighted criteria:

CriterionWhat the judges consider
Technological implementationQuality of engineering and effectiveness of Nebius/NVIDIA integration
DesignCompleteness and coherence of the product experience
Potential impactA credible, specific problem and real audience
Quality of the ideaCreativity and a genuine understanding of the problem space

Equal weighting makes balance important. A powerful orchestration stack is insufficient if users cannot understand the interface; a polished mock-up is insufficient if the required model integration does not function. Tie-breaking uses the criteria in listed order, followed by a judges' vote if necessary.

8. Prize breakdown: cash, hardware and bonus awards

AwardPrizeQuantity
Grand Prize$20,0001
Second place$10,0001
Third place$6,0001
Winner of each of the four tracksNVIDIA Jetson Orin Nano4 total
Best Use of Tavily$3,0001
City Winner$50020
Most Valuable Feedback$100 and NVIDIA swag pack10

The City Winner awards are linked to attendance at specified in-person Builders & Brews city events; online registration alone does not establish eligibility. The Tavily award requires a functional runtime Tavily API call. Feedback awards have their own submission condition.

Prize stacking is restricted: the rules say each project is eligible for one Overall Award or one Track Award, plus one Bonus Award. Awards remain subject to verification, forms, taxes and applicable fees. For teams, the designated representative handles allocation of any prize. The rules envisage delivery after completion of the required documentation, not instant payment upon an announcement.

9. Suggested development schedule before October 30

With the deadline close, the most useful strategy is to optimise for verifiable results. The following is a suggested engineering plan, not an official schedule:

  1. Define scope: one audience, one core task, one primary success metric and one track.
  2. Verify runtime integration: make an actual inference call to Token Factory or deploy via the listed Nebius AI Cloud route, with an NVIDIA open-source model.
  3. Complete a vertical slice: data in, model action, validation and output to the user.
  4. Test failures: timeouts, absent credentials, bad input, misleading output and permission boundaries.
  5. Prepare repo and README: include licence, assets, installation and testing steps.
  6. Record the demo: under three minutes, publicly available on YouTube; follow the one-minute footage condition for Physical AI.
  7. Submit and verify: open the Devpost submission confirmation rather than relying on a saved draft.

10. A useful project architecture

Browser / CLI / Device
        |
Application API
        |
Input checks and permission boundaries
        |
Nebius Token Factory API
        |
NVIDIA open-source model
        |
Output validation / optional tools
        |
User-visible result + evaluation log

Alternative compliant execution path:
Nebius AI Cloud compute deployment
(Serverless Jobs, Endpoints or DevPods),
with the required NVIDIA model.

This is a conceptual design sketch, not a copy-paste production manifest. Your actual integration must follow the provider's current authentication, model availability, limits and service documentation. Never put production API keys into a public repository or client-side application.

11. Existing products, multiple submissions and teams

Previously existing work is allowed if it receives significant new development after the start of the submission period, and the changes are explained. Multiple submissions are also permitted, but must be unique and substantially different; an interface colour change does not create a distinct product.

A team representative must be properly authorised. Document ownership and use of libraries, assets and datasets. If your employer or client owns relevant code, secure explicit permission before publishing it under an open-source licence. The rules also prohibit projects developed using certain prior financial or preferential support from the sponsor or administrator and allow exclusion for apparent conflicts of interest.

12. Intellectual property and publicity: read before publishing

The rules say that submissions remain the intellectual property of their creators and grant the sponsor a non-exclusive licence to use entries for judging. They also allow the sponsor and Devpost to publicise the submission and use contributor names, images, voices and likenesses in hackathon promotional materials during the event and for three years afterward, subject to the stated terms.

This does not override the separate obligation to release the project repository as open source. That is the practical trade-off to evaluate before entering with proprietary work. Publishing an open-source licence grants rights to code recipients under that licence, while the organisers' judging and publicity provisions address other uses.

Participants also accept release, indemnification, liability and dispute terms. The official rules contain arbitration provisions and New York governing-law language, subject to legal restrictions in relevant jurisdictions. Tax and banking obligations may apply to awards. This guide is not legal advice.

13. Common reasons a promising submission could fall short

  • Connecting to an unrelated API while only mentioning NVIDIA in the presentation.
  • Publishing a repository without a clear open-source licence or working setup instructions.
  • Uploading a video longer than the stated viewing window or keeping it unlisted from public access where visibility is required.
  • Submitting an old product without explaining significant work performed during the competition period.
  • Choosing the Physical AI track but omitting the specifically required one-minute operational demonstration.
  • Using private data, music or third-party assets without the rights to publish them.
  • Not providing the access judges need to test the project.
  • Saving a Devpost draft without completing the actual entry before the deadline.

14. Final submission checklist

  • Eligible entrant and authorised team representative, if applicable.
  • One of the four tracks selected.
  • Working application using Nebius Token Factory or Nebius AI Cloud and at least one NVIDIA open-source model.
  • Public and correctly licensed source repository with working README.
  • Working demo URL or applicable Physical AI procedure.
  • Public YouTube demonstration shorter than three minutes.
  • Explanation of the required models and Nebius components.
  • English-language materials or complete English translations.
  • Disclosure of substantial changes if the project existed before August 26.
  • Required tool feedback completed.
  • Final submission completed before October 30, 2026, 10:00 a.m. Pacific Time.

Conclusion

This hackathon rewards more than a clever model prompt. It asks for a working, open-source application, genuine Nebius and NVIDIA model usage, clear instructions and a focused demonstration. Builders who account for the submission requirements early can spend their last days improving the product rather than fixing avoidable compliance mistakes.

For the latest eligibility language, deadlines, award conditions and changes to the event, read the official sources below. The rules, rather than third-party guides, control in the event of a discrepancy.

Official references