Global Chess Challenge 2025
Challenge Rules
Global Chess Challenge 2025 — Official Rules
NO PURCHASE NECESSARY TO PARTICIPATE. VOID WHERE PROHIBITED.
By participating in this Challenge, you agree to abide by these Official Rules and all decisions of the Organizers.
Definitions
| Term | Definition |
|---|---|
| Sponsor | Amazon Web Services, Inc. ("AWS"), providing financial and brand sponsorship. AWS is not responsible for the organization or execution of the Challenge. |
| Organizer | AGI House, responsible for administering, managing, and operating the Challenge. |
| Platform Provider | AIcrowd SA, responsible for providing the online platform and leaderboard infrastructure. AIcrowd is not an Organizer or Sponsor. |
Overview
The Global Chess Challenge 2025 ("Challenge") is a global hybrid competition organized by AGI House and sponsored by Amazon Web Services (AWS). Infrastructure and leaderboard services are provided by AIcrowd.
The competition focuses on leveraging AWS Trainium chips to train and evaluate a family of Small Language Models (SLMs) specialized for reasoning in text-based chess environments. Participants are tasked with building text-only models that, given a chess board position, can:
- Produce legal chess moves
- Generate succinct textual explanations justifying those moves
- Operate efficiently within the computational constraints of AWS Trainium hardware
All submitted models are executed independently by the Organizers on controlled infrastructure. Submissions must operate as standalone language models, producing moves solely via token prediction. External tool use, heuristic search procedures, or runtime access to chess engines or auxiliary decision systems is not permitted.
The challenge emphasizes reasoning efficiency, strategic quality, and hardware-aware optimization, rather than raw model scale. Stockfish serves as the objective verifier for move legality and strategic soundness, while large public chess corpora may be used to provide scale for training and evaluation.
Sponsor
Amazon Web Services, Inc., 410 Terry Ave North, Seattle, WA 98109. AWS is the sponsor of the Challenge and is not responsible for the administration or organization of the event.
Organizers & Admins
AGI House is the sole Organizer of the AWS Trainium Challenge 2025. "Organizer Admins" are companies or entities authorized by the Organizer to support execution of the Challenge. For this event, this includes AIcrowd SA as a platform and evaluation infrastructure provider. Organizer Admins are not responsible for the overall organization or legal administration of the Challenge.
Timeline, Website & Registration
| Milestone | Date |
|---|---|
| Launch & Registration Opens | AWS re:Invent, December 4, 2025 |
| Round 1 Submissions Close | December 31, 2025 (23:55 UTC) |
| Round 2 Submissions Close | January 31, 2026 (23:55 UTC) |
| Final Tournament | February 1–7, 2026 |
| Winners Announced | February 15, 2026 |
Challenge Evaluation
Overview
Evaluation proceeds in multiple stages designed to assess move quality, reasoning consistency, and competitive strength across both baseline play and peer-to-peer competition.
Round 1 & Round 2: Baseline Evaluation
All submitted models are evaluated against fixed Stockfish opponents to establish a stable, comparable performance baseline. For each round:
- Each submission plays 50 games against Stockfish Skill Level 0 (Depth 1) and 50 games against Stockfish Skill Level 0 (Depth 5)
- Game positions, time controls, and computational resources are standardized across all submissions
Scoring:
| Metric | Description | Priority |
|---|---|---|
| Average Centipawn Loss (ACPL) | Computed by analyzing each game using Stockfish Level 20 (Depth 20). Lower ACPL = stronger play. | Primary |
| Win Rate | Calculated across all games played against both Stockfish configurations. | Secondary (tie-breaking) |
Important: The reasoning text is not scored directly. Evaluation is based exclusively on the UCI move provided within the
<move>tags. Any text outside these tags is ignored.
If a submission fails to provide a valid UCI move within the <move> tags, the evaluation system will retry up to three (3) times. If still invalid after retries, the model is treated as having resigned and the game is recorded as a loss.
Eligibility for Final Tournament
At the conclusion of Round 2, submissions that achieve an ACPL lower than the official baseline model defined by the Organizers will advance to the Final Tournament.
Final Tournament: Swiss-Style Competition
Eligible submissions compete in a Swiss-style tournament to determine final placements and prize winners.
- All eligible models participate in a fixed number of rounds (determined by the number of qualifying submissions)
- Pairings use a Swiss-system format, ensuring models with similar records compete against each other
- Each match consists of a predefined number of games with alternating colors and standardized time controls
- ACPL is not used in the Final Tournament — only win/loss/draw results count
Tournament Scoring
| Result | Points |
|---|---|
| Win | 1 |
| Draw | 0.5 |
| Loss | 0 |
Final rankings are determined by total tournament points accumulated across all Swiss rounds.
Tie-Breaking Rules
If two or more submissions finish with equal total points, tie-breaks are applied in this order:
- Head-to-head result(s) between tied submissions, if applicable
- Buchholz score (sum of opponents' final tournament points), or an equivalent strength-of-opposition metric
- Sonneborn–Berger score, where applicable
- Any additional tie-breaking procedure announced by the Organizers prior to the Final Tournament
No engine-quality metrics (including ACPL) are used during the Swiss-style Final Tournament. Only game outcomes are considered.
Eligibility
You are eligible if you (and each team member):
- Are an individual
- Are 18 years or older (and of majority age in your jurisdiction)
- Have internet access, an email account, and a personal computer
Residents of the following regions are not eligible for prizes (including cash prizes or compute credits):
Crimea region of Ukraine, Cuba, Iran, North Korea, Sudan, Syria, Quebec (Canada), Brazil, Italy, Russian Federation.
Note: Residents of these regions may still participate and retain their leaderboard rank. Any cash prizes will be passed to the next eligible team.
Employees, contractors, or household members of Amazon, AGI House, and AIcrowd are not eligible to receive prizes.
You are responsible for reviewing and understanding your employer's and country's policies regarding participation. If you participate in violation of these policies, you and your Entry may be disqualified.
If your employer has rights to intellectual property created during employment, you must confirm you have received authorization to participate, or that your employer has waived such rights for this Challenge. Organizers may request written confirmation.
Entry Requirements
The Entry MUST:
- Be compatible with the official submission format
- Be in English
- Be the Team's own original work
- Not have been submitted previously in any promotion of any kind
- Not contain inappropriate, offensive, defamatory, or illegal content, or content that promotes discrimination, violence, or infringes on others' rights
Team Members MUST:
- Obtain all necessary consents, approvals, and licenses for submission
- Obtain agreement from all team members to these Rules
- Not engage in false, fraudulent, or deceptive acts
- Not tamper with or abuse any aspect of the Challenge
- Ensure transparency in the use of external data sources with clear documentation
- Ensure external datasets are permissible for non-commercial or academic research and comply with licensing terms
- Provide rationale for any declared datasets
Violation of these terms will result in immediate disqualification.
Submission Guidelines
Timeouts and Infrastructure Requirements
Submissions are evaluated on a standard hardware configuration disclosed prior to the deadline. Each submission is subject to strict runtime limits per task, including per-turn timeouts. Submissions that exceed time limits or fail to execute on the designated infrastructure may be marked as failed.
Participants are responsible for ensuring compatibility with evaluation hardware and accounting for constraints such as memory limits and disabled external network access.
Submission Limits and Documentation
- Teams may submit up to 20 entries per day
- Winning solutions must be validated for compliance with allowed data and thoroughly documented
- Submissions must achieve an ACPL lower than the official baseline model to qualify for prizes
- Winners must submit a comprehensive solution report for publication on a platform designated by the Organizers
Eligible Models and Execution Backends
Only a subset of models and execution backends are eligible, reflecting currently supported and performance-optimized configurations for AWS Trainium. Participants must use only those model architectures and inference/training backends explicitly supported by the official Challenge infrastructure:
Supported Model Types & Backends →
Model Execution Constraints
All decisions must be produced solely through token-level language model inference. The following are prohibited at inference time:
- External tools or function calling
- Heuristic search procedures
- Runtime interaction with chess engines or auxiliary decision systems
- Any form of external computation, search, retrieval, or handcrafted rule-based logic
- Embedded engines or hidden heuristics
Submissions that attempt to bypass these constraints may be disqualified at the Organizers' discretion.
Model Size Limitation
Only models with a total parameter count of strictly fewer than 8,500,000,000 (8.5B) parameters are eligible for participation, leaderboard ranking, Final Tournament qualification, and prizes.
The parameter count is determined by the total number of trainable parameters at inference time, excluding optimizer states but including all model weights. Submissions exceeding this limit may be disqualified.
Disqualification
The Organizer reserves the right to disqualify any Entry and/or participant if:
- Any team member or the Entry is found to be ineligible
- A participant creates or uses multiple accounts
- There is fraud, deception, or interference with the leaderboard
- Any activity that undermines the integrity of the Challenge is detected
Violating the multiple-accounts rule will result in immediate disqualification and potential exclusion from future events.
Intellectual Property and Use of Entries
Participants retain ownership of their Entries, including any models or code developed during the Challenge. By submitting an Entry, participants grant AGI House, AWS, AIcrowd SA, and their respective partners, sponsors, and affiliates a non-exclusive, worldwide, royalty-free, irrevocable license to:
- Review, evaluate, and display the Entry in connection with the Challenge
- Use the Entry for research, evaluation, publication, and promotional purposes
- Use the Entry in any media for non-commercial or commercial purposes related to marketing or promotion of Organizers, sponsors, and their products
- Publish results, model performance data, outputs, evaluation results, and descriptions of winning approaches
The Base Model and any datasets provided by Organizers remain their exclusive property and may only be used within the scope of the Challenge. Trained or derivative models built upon the Base Model must remain open-sourced under a permissive license (e.g., MIT, Apache 2.0).
Participants acknowledge that:
- They will not be compensated and may not be credited (at Organizers' discretion) for use of the Entry
- Organizers, sponsors, and partners may have developed similar materials
- Entries may be posted publicly and shared with media partners and third parties
- Organizers are not obligated to use the Entry in any way, even if selected as a winner
Winner Selection and Notification
Entries are ranked on the AIcrowd leaderboard based on algorithmic scoring (Primary: ACPL; Secondary: Win Rate) during Round 1 and Round 2. Final winners are determined by the Swiss-style Final Tournament using game outcomes only.
Potential winners are contacted via the email address associated with the Team Leader's AIcrowd account. In the case of a single-participant Entry, that individual is deemed the Team Leader.
If a potential winner cannot be contacted, does not respond, refuses the prize, or is found ineligible, the prize may be awarded to the next eligible Team on the leaderboard.
Solution Report Requirement
As a condition of prize eligibility and acceptance, winners must submit a comprehensive solution report prepared for publication on a platform designated by the Organizers. Reports must accurately describe the technical approach used. Submissions that are materially incomplete, misleading, or consist primarily of generic content without meaningful technical detail may result in disqualification or forfeiture of prizes.
Prizes and Conditions
Prize amounts and award categories will be announced on the official Challenge webpage.
- Prizes are non-transferable and may not be substituted except at the Organizer's discretion
- Winners are solely responsible for all applicable taxes, withholdings, and costs associated with prizes
- Winners must complete required documentation (e.g., W-9 or W-8BEN tax forms) within 14 days of notification
- Failure to provide complete documentation may result in forfeiture
- The Organizer reserves the right not to award a prize if a winning Entry cannot be reproduced, verified, or validated
- Prizes will be awarded within 6 months from the end of the Challenge
- Disputes among team members regarding prize sharing are not the responsibility of Organizers
A list of winners will be posted on AIcrowd and may be announced on official AWS or AGI House channels.
Governing Law and Dispute Resolution
This Challenge and any dispute arising under or related to it will be governed by the laws of the State of California, without regard to conflict of law principles. All claims and disputes must be resolved exclusively in the state or federal courts located in San Mateo County, California.
AWS, as Sponsor, will not serve as a party to dispute resolution.
Participants waive any right to participate in a class action or class arbitration related to the Challenge.
Personal Data and Privacy
Organizers may use cookies and/or collect IP addresses for implementing rights and obligations under these Rules. Personal data may be used to:
- Contact you regarding the Challenge
- Confirm details of your Entry
- Administer and execute the Challenge, including sharing with Organizer Admins and external partners for prize distribution
- Credit you and/or your team or identify you as a Winner
- Meet obligations under these Rules or applicable law
Only name and email address are required to participate. Please read the AIcrowd Site terms and conditions carefully to understand how your data may be used by AIcrowd SA.
Force Majeure and Event Disruption
Organizers reserve the right to cancel, suspend, or modify the Challenge if circumstances beyond their reasonable control compromise the fairness, integrity, or feasibility of the competition — including natural disasters, war, public health emergencies, system failures, or regulatory actions.
Organizers may make reasonable updates to the structure, schedule, evaluation criteria, and technical framework of the Challenge. Updates will be communicated via official Challenge channels. Participants will not be required to formally re-accept these Rules unless updates materially alter eligibility, prize structure, data use rights, or other legally significant terms.
Limitation of Liability
To the fullest extent permitted by law, you agree that the Organizer and their directors, officers, employees, agents, and assigns will not be liable for personal injuries, death, damages, expenses, costs, or losses of any kind resulting from:
- Participation or inability to participate in this Challenge
- Acceptance, use, or inability to use a prize
If any portion of this Challenge is compromised by virus, bugs, unauthorized intervention, or other causes beyond Organizers' control, Organizers may: (a) cancel the Challenge; (b) pause the Challenge until issues are resolved; or (c) consider only Entries submitted prior to compromise.
If any provision of these Rules is deemed unenforceable, the remaining provisions remain valid and in effect.