Paradigm

The Muse Spark Open Kernel Challenge is a competition presented by Meta, Paradigm, and Andromeda to improve the speed of Meta’s Muse Spark Open model. It invites developers to optimize GPU kernels and, in a separate serving track, modify SGLang to serve the model more efficiently. Participants can use tools such as Triton, CUTLASS, FlashInfer, and DeepGEMM, or write custom CUDA kernels.

The challenge provides a command-line tool for installation, sign-in, initializing problems, submitting code, and viewing leaderboard results. Kernel submissions are tested for correctness on previously unseen inputs and benchmarked against SGLang on an H200 GPU. The serving track evaluates modified SGLang on an eight-H200 node using multi-turn coding conversations and varying numbers of concurrent users.

Performance is assessed through speed and throughput measures. The competition offers cash prizes, including $50,000 for first place, and top contestants may have an opportunity to interview for a role at Meta. The site also features challenge rules, problem details, leaderboards, and information about the prize schedule