AWS Trainium Frontier Competition

AWS Trainium Frontier Competition Terms and Conditions

DESCRIPTION OF CONTEST:

From August 6, 2026 11:45pm PST to September 30, 2026 11:45pm PST be one of the first 100 teams to register to compete in the AWS Trainium Frontier Competition for a chance to win up to $25,000 USD. NO PURCHASE NECESSARY. VOID WHERE PROHIBITED. Terms and conditions apply.

CONTEST TERMS AND CONDITIONS:

By competing in the AWS Trainium Frontier Competition (“Contest”), you agree to be bound by these Terms and Conditions.  NO PURCHASE NECESSARY.  VOID WHERE PROHIBITED.

From August 6, 2026 11:45pm PST to September 30, 2026 11:45pm PST (“Registration Period”) be one of the first 100 teams to register to compete in the Contest. From August 31, 2026 to November 4, 2026 (“Contest Period”), compete for a chance to win up to $25,000 USD.

Participants must have registered for the Contest during the Registration Period, and be one of the first 100 teams to register for the Contest. Must be at least 18 years of age. Excludes individuals living in Argentina, Australia, Brazil, France, Hong Kong, Indonesia, Italy, Malaysia, Poland, Philippines, Thailand, Vietnam, Singapore, Russia, Spain, Cuba, Iran, Syria, North Korea, Belarus, the region of Crimea, and the United Arab Emirates. Employees, interns, and scholars of Amazon Web Services (“AWS”) and its affiliates in either 2025 or 2026 and their immediate family members are ineligible to participate. 

Teams of eligible individuals may have a maximum of four (4) members (“Teams”). By entering a submission on behalf of a Team, you represent and warrant that you are the representative authorized to act on behalf of your Team. Limit one entry per Team.

To compete, Teams will use the provided Contest starter kit to develop custom kernels using the Neuron Kernel Interface (NKI) for the Qwen3-30B-A3B Mixture of Experts model and optimize inference performance on AWS Trainium hardware. Academic Teams competing in the Contest will be provided with AWS Promotional Credits after registration for development and testing. "Academic Team" means a Team composed primarily of students enrolled at an accredited academic institution. Such AWS Promotional Credits will expire one (1) calendar year from the date of issuance. AWS Promotional Credits are subject to the AWS Promotional Credits terms and conditions found at: https://aws.amazon.com/awscredits/. All Contest materials, rules, and updates are published onhttps://github.com/aws-neuron/trainium-frontier#

Round 1

Teams will compete in Round 1 from August 31, 2026 to September 30, 2026, and must make a submission that trains a model on Trn2 per the specifications of the competition repository to a leaderboard as designated by AWS (“Leaderboard”) by 11:59 PM PT on September 30, 2026 for consideration. The top 10 Teams from Round 1 (the “Round 1 Winners”) will be selected based on Round 1 Evaluation Criteria below. Round 1 Winners will be notified by 11:59 PM PT on October 7, 2026, of their acceptance into Round 2 using the email address provided during registration.

Round 1 Evaluation Criteria

Submissions must pass a correctness gate before becoming eligible for evaluation. To pass the correctness gate, submissions must: (1) complete training within the applicable time budget (30 minutes), (2) produce a valid score (no NaN, no crashes), (3) run on the provided hardware and software environment without modifications to the evaluation harness, (4) use the full chip, and (5) use only the provided training data without pre-trained weights, external data, or CORE eval bundle content in training.

Round 1 Scoring: Pure val_bpb (validation bits per byte)

·       Scoring metric: val_bpb (validation bits per byte) on a pinned validation shard, computed over ~20M tokens. Lower is better.

·       Training budget: 30 minutes of wall-clock training on a single Trn2 chip (trn2.3xlarge), excluding startup/compilation.

·       Advancement: Top 10 Teams by lowest val_bpb advance to Phase 2.

Tie Breaker: In the event of a tie, Round 1 Winners will be selected based upon the following criteria:

· Fewer total parameters

· Simpler code

Round 2

Round 1 Winners will compete in Round 2 from October 7, 2026 to November 4, 2026, and make their submission to the Leaderboard by 11:59 PM PT on November 4, 2026 for consideration. In Round 2, Teams are asked to make a submission that trains a model on Trn2 and performs inference, each per the specifications of the competition repository. Round 2 submissions are scored on a weighted combination of a training-efficiency metric and a capability metric as provided in the Round 2 Evaluation Criteria below.

Round 2 Evaluation Criteria

Submissions must pass a correctness gate before becoming eligible for evaluation. To pass the correctness gate, submissions must: (1) complete training within the applicable time budget (4 hours), (2) produce a valid score (no NaN, no crashes), (3) run on the provided hardware and software environment without modifications to the evaluation harness, (4) use the full server, and (5) use only the provided training data without pre-trained weights, external data, or CORE eval bundle content in training. In Round 2, submissions must produce an inference path consistent with the trained model and complete the CORE evaluation pass within the 60-minute scoring timeout.

Round 2 Scoring: Weighted val_bpb + CORE (50% / 50%)

·       Training metric (50%) — val_bpb: validation bits per byte on the pinned validation shard, computed on the final trained model. Lower is better.

·       Capability metric (50%) — CORE: the CORE metric as defined in the nanochat evaluation suite. CORE is an aggregate score across a pinned battery of in-context learning tasks spanning reading comprehension, commonsense reasoning, language understanding, world knowledge, and symbolic problem-solving.

For each task, raw accuracy is centered against the task's random baseline. The centered score for a task is the model's accuracy on that task, rescaled so that a model performing at the task's random baseline scores 0 and a model answering every question correctly scores 1. The final CORE metric is the arithmetic mean of centered scores across all tasks in the eval bundle.

CORE is evaluated in scoring mode (log-likelihood over candidate completions, no autoregressive generation), using the nanochat core_eval harness against the full pinned eval bundle. Higher is better.

Final score: Both metrics are min-max normalized across the Round 2 submission pool, directions aligned (CORE flipped so lower is better), then averaged with 50/50 weighting. The final leaderboard is ranked by this composite score.

Tie Breaker: In the event of a tie, Finalists will be selected based upon the following criteria:

· Fewer total parameters

· Simpler code

The top three (3) Teams who receive the most points from Round 2 will be designated as the “Finalists”. Finalists will be notified by 11:59 PM PT on November 11, 2026, using the email address provided during registration.

Final Round

Finalists will have an opportunity to present their solutions at the NeurIPS 2026 Competition Event (the “Conference”), December 6–12, 2026. Finalists must each have at least one (1) team member available to present in person at the Conference. Attendance at the Conference, including but not limited to travel, lodging, meals, and any other expenses, is at each Team's own cost and expense. Finalists must provide the additional deliverables, as outlined below, by the close of the NeurIPS 2026 Conference on December 12, 2026. Failure to complete the Finalist Deliverables by the required deadline may result in disqualification at AWS's sole discretion.

Finalists Deliverables

This Contest requires the use and development of open-source materials. Open source requirements are as follows:

Prizes

·       Winners will receive prizes as follows. 

o   First Place: $25,000 USD

o   Second Place: $10,000 USD

o   Third Place: $5,000 USD

·       Finalists: One month of dedicated single-chip Trn3 access for research use after the competition concludes

Entries that do not meet the requirements set forth in these Terms and Conditions may be disqualified in AWS’s sole discretion. Approximate retail value of the first place prize is $25,000 USD, approximate retail value of the second place prize is $10,000 USD, and the approximate retail value of the third place prize is $5,000 USD. Prizes will be distributed equally among team members. AWS reserves the right to substitute a prize (or portion thereof) for an item of comparable or greater value, at AWS's sole discretion. Eligible teams will receive the prize in accordance with the instructions provided to you by Devpost within 60 days following the conclusion of the Conference.

If you are selected as a potential winner of the Contest, AWS or its representative will contact you using the email address provided during registration.  Failure to provide the requested information or to respond to communications about the Contest within a reasonable period of time, as determined by AWS in its sole discretion, may result in the forfeiture of the prize.

By accepting the prize, you confirm that your receipt is neither prohibited nor inconsistent with any applicable laws, regulations, or binding orders, including applicable ethics or procurement rules, your receipt will not create a conflict of interest for AWS, and there are no ongoing competitive procurements for which your receipt of this benefit could conflict AWS from participating in the Contest.

The following personal information will be collected by AWS for the Contest solely for the purposes of administering the Contest and verifying participant eligibility: name, city, state or province, country.  Failure to provide all necessary personal information may result in participants’ participation in the Contest being deemed ineligible.  AWS handles your information in accordance with AWS Privacy Notice.

A winners list will be publicly available on the Contest GitHub repository  for up to one year after December 12, 2026.  Acceptance of prize by the winner constitutes permission for AWS to use winners’ names or likenesses, and city, state or province, and country, if submitted to AWS, for any disclosures required by law, including a winners list, and for advertising and promotional purposes relating to the Contest in any and all media now or hereafter devised, worldwide in perpetuity (or to the maximum extent permissible under applicable law), without additional compensation, notification or permission, unless prohibited by law.

By submitting your submission to the Contest, you represent and warrant that your submission does not: (i) perform or promote activities that are illegal; (ii) violate or infringe the intellectual property, proprietary, or other rights of others; (iii) perform or promote activities that are offensive or disparaging in any manner; (iv) for Academic Teams, violate your institutional academic integrity policies, or (v) cause harm to others or to AWS’s operations or reputation.

By submitting your kernel to the Contest, you represent and warrant that you have all rights, permissions, and authority necessary to grant AWS the right to use, reproduce, display, publish, and otherwise publicly disclose submission materials, including for promotional, marketing, and informational purposes, without further consent, approval, or compensation.

The Contest is sponsored by AWS. AWS reserves the right to cancel or modify the Contest at any time and for any reason. AWS and its affiliates and licensors will not be liable to you under any cause of action or theory of liability relating to the Contest, even if a party has been advised of the possibility of such damages, for (a) indirect, incidental, special, consequential, or exemplary damages, or (b) lost profits, revenues, customers, opportunities, or goodwill. In any case, AWS and its affiliates’ and licensors’ aggregate liability under these Terms and Conditions will not exceed USD $100. The limitations in this paragraph shall apply only to the maximum extent permitted by applicable law.