A $100,000 challenge aims to speed up Meta’s Muse Spark Open

Daniel Gross announced a $100,000 challenge on Oct. 22, 2026, to make Muse Spark run faster. The contest page asks entrants to speed up Meta’s Muse Spark Open in two ways: by optimizing its GPU kernels and by changing how SGLang serves the model.

Two routes, different tests

For kernel submissions, entrants write code in a submission file. The contest rules say they can use components SGLang uses or write their own CUDA. Each kernel must match a reference implementation within the problem’s tolerance on random inputs it has not seen. For timing, the entrant’s kernel and SGLang’s take turns with the same inputs on one H200.

Serving submissions instead consist of a diff of changes to SGLang. According to the page’s workload description, the modified code serves Muse Spark Open on a full eight-H200 node using hidden multi-turn coding conversations at one, four, 16 and 32 concurrent users. Its output must agree with stock SGLang nearly as closely as stock SGLang agrees with itself. The page says the score compares throughput (tokens/s per GPU) and interactivity (tokens/s per user) against SGLang across those loads.