FlagOS Open Computing Global Challenge – Track 2 is a multi-season, comprehensive competition jointly hosted by the FlagOS Community and the IEEE. Season 2 officially opens for registration on July 19, 2026, and runs from July to December 2026. It is co-hosted by the FlagOS Community, BAAI, and IEEE, with leading model provider OpenBMB and state-of-the-art inference framework SGLang as co-organizers.
This track focuses on Inference Throughput Performance Optimization for MiniCPM5-2B. Following the large-scale deployment of trillion-parameter models, high computational and video-memory overhead has led to substantial inference latency and low throughput. This competition addresses these industry pain points by challenging participants to achieve efficient distributed end-to-end inference with fixed hardware resources while guaranteeing accuracy and stability.
Participants are expected to leverage full-stack techniques including optimized model parallelism strategies, video-memory management (KV Cache optimization, dynamic batching), computation logic optimization (operator fusion, kernel rewriting), and communication scheduling optimization. Based on the FlagOS inference framework and the high-performance operator library FlagGems, contestants must fully tap the maximum performance potential of the MiniCPM5-2B model developed by FaceWall under the premise of strictly preserving model accuracy without materially increasing latency.
Eligibility and Participation: All individuals or teams are welcome to register, with no restrictions on nationality, age, or occupation. Participants may compete individually or as a team (maximum 5 members per team). Each individual or team may only hold one participant identity; duplicate registrations are not permitted.
Awards: The total prize pool for Track 2 is $9,000: * First Prize: $4,500 * Second Prize: $3,000 * Third Prize: $1,500
Detailed competition rules will be released in mid-September.