FlagOS Open Computing Global Challenge

Overview The "FlagOS Open Computing Global Challenge" is a multi-season competition jointly hosted by FlagOS Community, Beijing Academy of Artificial Intelligence (BAAI) and CCF Open Source Development Technology Committee (ODTC). By encouraging developers to engage in hands-on creation and innovative exploration based on the unified AI system software stack provided by FlagOS, the competion aims to enhance the skills of AI developers. With a total prize pool of 2 million RMB, winners will share the rewards and receive official honorary certificates from the organizers. We cordially invite AI algorithm and system engineers, hardware and compiler developers, as well as all technical practitioners and visionaries dedicated to overcoming the complexities of AI chip programming to join this challenge and collectively advance the thriving development of open computing ecosystem. Season 1 will open for registration on January 9, 2026, with final results to be announced in early June 2026. This season focuses on three core tracks: Operator Development and Optimization, LLM Inference Optimization, and Automatic Data Annotation. The goal is to deeply optimize the performance and operational efficiency of large language models while driving technical implementation and industrial innovation. S1 Challenge Overview This season features three distinct tracks, each targeting specific dimensions of development expertise: Track 1: Focuses on low-level operator implementation and cross-hardware deep optimization capabilities. Track 2: Focuses on the maximal excavation of full-stack inference performance and system-level tuning expertise. Track 3: Focuses on the ability to perform automated data annotation within long-context scenarios. Track 1: Operator Development and Performance Optimization Based on FlagOS This track is designed around the practical application requirements of FlagGems, the universal operator library within FlagOS ecosystem. It comprises 20 operator development tasks that primarily evaluate participants' operator development and extreme performance optimization capabilities. The organizing committee will conduct quantitative scoring of submissions based on six dimensions: Functional Correctness, Performance Competitiveness, Open-Source Adaptability, Cross-hardware Compatibility, Test Case Completeness, Code Readability. Task Difficulty Distribution: DifficultyNumberDescriptionEasy8 tasksBasic mathematics / General fundamental operatorsMedium8 tasksClassic deep learning operatorsHard4 tasksCutting-edge complex scenario operators Track 2: Large Model Inference Throughput Optimization Based on FlagOS Multi-Chip Framework This track focuses on evaluating participants' ability to optimize large model inference performance. Participants are required to perform extreme inference throughput optimization on Qwen3-4B while strictly maintaining model accuracy and without significantly increasing latency. Participants must comprehensively utilize parallel strategies, memory management, compute kernels, and sampling algorithms to unlock maximum performance potential on the designated framework and hardware. Designated Resources: Large Model Inference Framework: vllm-plugin-FL, High-Performance Operator Library: FlagGems, Model: Qwen3-4B. Track 3: Automatic Data Annotation for Large Models in Long-Context Scenarios This track focuses on evaluating participants' model application capabilities. Each participant or team is required to design an effective In-Context Learning (ICL) solution for automatic data annotation based on Qwen3-4B model, using the unified dataset provided by the organizing committee, and perform inference validation on the standardized evaluation dataset. The organizing committee will compile evaluation scores and rankings for all participants in a standardized leaderboard format. Prize Structure Track 1 Track 2 Track 3 Easy Operator Task: 1000 RMB per task 1st Prize: 30,000 RMB 1st Prize: 30,000 RMB Medium Operator Task: 2000 RMB per task 2nd Prize: 20,000 RMB 2nd Prize: 20,000 RMB Hard Operator Task: 3000 RMB per task 3rd Prize: 10,000 RMB 3rd Prize: 10,000 RMB Competition Schedule Phase Track 1 Track 2 Track 3 Registration Jan 9 – May 20 Jan 9 – May 20 Jan 9 – May 20 Computing Resource Application — Jan 9 – Mar 11 Jan 9 – Mar 11 Computing Resource Review & Allocation — Mar 12 – Mar 23 Mar 12 – Mar 23 Development Feb 24 – May 20 Mar 24 – May 20 Mar 24 – May 20 Submission of Technical Reports and Source Code — May 21 – May 31 Evaluation May 21 – Jun 4 May 21 – Jun 4 May 21 – Jun 4 Results Announcement Early June Early June Early June Rules and Guidelines 1. Eligibility and Participation Eligible Participants: All individuals or teams interested in the competition are welcome to register. Participation Format: Participants may compete individually or as a team, with a maximum of 3 members per team. For team registration, only one representative needs to submit information about your team and your works — please avoid duplicate submissions from multiple team members. Each individual or team may only hold one participant identity throughout the competition period; duplicate registrations are not permitted. Registration Requirement: All participants must complete the Registration Form before participating: Please ensure all registration information is truthful, accurate, and valid. Failure to comply may result in disqualification and forfeiture of rewards. 2. Track Selection Rules Each participant (individual or team) may only submit work to one track. Once a submission is made to any track, the track selection is considered confirmed and cannot be changed. Duplicate submissions across different tracks will be deemed invalid, and the organizers reserve the right to disqualify such participants. 3. Submission Requirements Content Standards: All submissions must comply with universally recognized international standards and public order and good morals. Content involving security threats, pornography, ethnic discrimination, religious discrimination, or privacy violations is strictly prohibited. The organizers reserve the right to identify and address any content deemed inappropriate for public display or distribution in global settings. Any harassment, discrimination, or other inappropriate behavior will result in disqualification. The FlagOS Community reserves the right to disqualify any participant or team at its sole discretion. Originality and Copyright: All submissions must be original work created by the participant (individual or team) with complete intellectual property rights. Plagiarism, copying, or unauthorized use of others' work or ideas is strictly prohibited. Verified violations will result in immediate disqualification and forfeiture of any awards, with the participant bearing full legal responsibility. If third-party materials (such as open-source code) are used, participants must clearly indicate the source upon submission and ensure proper authorization has been obtained without infringing on any third party's legal rights. Computing Resource Usage: Computing resources provided by the organizing committee are exclusively for competition-related experiments. Lending resources to others, cryptocurrency mining, or any tasks unrelated to the competition are strictly prohibited. Submission Format and Method: Please prepare your work strictly according to the official submission format requirements for each track and submit through the officially designated channel before the submission deadline. Late submissions or submissions that do not comply with format requirements will not be eligible for evaluation. 4. Evaluation and Awards Evaluation will be conducted based on the published criteria for each track. Awards will be determined independently for each track based on participation. 5. Additional Notes By submitting work, participants agree that the organizers and their authorized entities have the right to use the work for promotion, exhibition, publication, and other purposes without compensation. Participants must ensure the authenticity of all submitted information. The organizers reserve the right to disqualify participants who provide false information. The organizers hold final interpretation rights over all competition rules and arrangements. In case of any disputes, the organizers' interpretation shall prevail. Contact Information For any questions, please contact: contact@flagos.io