Hi everyone 
Thank you for the thoughtful questions and feedback over the past weeks. Based on those discussions, we’ve consolidated and clarified the Official Rules for the Global Chess Challenge 2025.
Action required: To continue participating, please visit the challenge page, click Participate, and accept the updated rules.
The challenge overview page has been updated to reflect the latest and most accurate information. Please review it for the newest details, links, and resources.
Timeline (UTC): Deadline Extension
- Round 1: Dec 4, 2025 → Dec 31, 2025 (23:55)
- Round 2: Jan 1, 2026 → Jan 31, 2026 (23:55)
- Final Tournament (no new submissions): Feb 1 → Feb 7, 2026
- Winners announced: Feb 15, 2026
Note: Model submissions close on January 31, 2026. From February 1 onward, no new submissions are accepted; this period is reserved for the Final Tournament of eligible models and post-challenge verification.
Evaluation Structure (Rounds 1 & 2)
Rounds 1 and 2 use a baseline evaluation against fixed Stockfish opponents to ensure stability and comparability.
Each submission plays:
- 50 games vs Stockfish Skill 0 (Depth 1)
- 50 games vs Stockfish Skill 0 (Depth 5)
All games use identical positions, time controls, and compute constraints.
All evaluations are performed on a standardized AWS Trainium configuration, specifically a trn1.2xlarge instance, to ensure consistency and fairness across all submissions.
Leaderboard Scoring
-
Primary metric: Average Centipawn Loss (ACPL)
- Computed using Stockfish Level 20 (Depth 20) as the reference evaluator
-
Secondary metric: Win Rate
- Used for tie-breaking and additional analysis
Important: Reasoning Text vs Scoring
- Submissions must output both:
- a chess move
- a short textual explanation
- Only the move inside
<uci_move>tags is scored - Reasoning text is not evaluated or scored
Invalid move handling
- Missing / malformed / illegal
<uci_move>→ up to 3 retries - Still invalid after retries → treated as a resignation (loss for that game, and a 1000 CPL for the resignation move)
Resignation & ACPL Handling (Clarification)
To address an issue identified with ACPL calculation for very short games, the evaluation logic has been updated as follows:
- Intentional resignations are permitted, but
- A resignation now incurs a fixed penalty of +1000 centipawns (CPL) applied to the resignation move.
This change ensures that:
- Models cannot artificially achieve low ACPL by resigning early, and
- ACPL remains comparable across games of different lengths.
This penalty applies only to the resignation move and does not otherwise alter move evaluation or game outcome handling.
Eligibility & Final Tournament
Advancement Criteria
- After Round 2, submissions with ACPL lower than the official baseline model become eligible for Finals.
♜ Final Tournament Format
- Swiss-style tournament
- Rankings based only on game outcomes:
- Win: 1 point
- Draw: 0.5 points
- Loss: 0 points
- ACPL is not used in Finals
Tie-breaks (in order)
- Head-to-head result (if applicable)
- Buchholz (or equivalent strength-of-opposition metric)
- Sonneborn–Berger (where applicable)
- Any additional rule announced before the Finals
♜ Final prize winners are determined exclusively by the results of the Swiss-style Final Tournament, in accordance with the tournament scoring and tie-breaking rules. ♜:trophy:
Submissions
- Submission cap: increased to up to 20 submissions per team per day
Eligible Models & Backends
Only officially supported model types and execution backends are eligible, and all evaluations are performed exclusively on AWS Trainium using the AWS Neuron + vLLM backend.
Supported list:
global-chess-challenge-2025-starter-kit/docs/neuron-and-vllm-tuning.md at master · AIcrowd/global-chess-challenge-2025-starter-kit · GitHub.
Note: We recognize that some teams are currently facing challenges submitting models targeting the AWS Neuron backend for execution on AWS Trainium.
The organizers are working closely with the AWS team to improve support responsiveness and unblock submission issues, so that all compliant models can be evaluated under the official infrastructure.
Execution Constraints
All models are executed on organizer-controlled infrastructure and must operate as standalone language models:
-
No tool calling -
No web access
-
No chess engines
-
No heuristic search or auxiliary decision systems
All decisions must be produced solely via token-level inference from the provided text input.
Attempts to bypass these constraints may result in disqualification.
Final Notes
Thanks again for the feedback and patience. We hope these clarifications make the evaluation and Finals structure fully transparent.
We’re excited to see how far participants can push reasoning quality, efficiency, and hardware-aware optimization on AWS Trainium.
All the best,
Team Global Chess Challenge
