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.
The listed prizes
The prize list offers $50,000 and a signed H200 for first place, $30,000 for second and $10,000 for third. It lists $3,000 for fourth, $2,000 for fifth and $1,000 each for sixth through 10th. The cash awards total $100,000, with the H200 listed alongside the first-place cash prize.
Comments