Muse Spark Open Kernel Challenge

We’re launching the Muse Spark Open Kernel Challenge in collaboration with Meta and Andromeda. Compete to make Muse Spark Open, Meta’s 816B-parameter mixture-of-experts model and the most capable American open-weight model to date, run faster on H200s. The fastest submissions win a share of $100,000 and first place also receives an NVIDIA H200 GPU signed by the Meta team.

A model’s weights are fixed when they are released. How useful it is afterwards depends on how fast and cheaply it can be served. This often involves a tradeoff: batching many user requests together makes each token cheaper, but makes each user wait longer.

Kernels are a big part of what decides this tradeoff. Better kernels allow for more users at the same speed, or faster responses at the same cost.

The Challenge

The challenge consists of two tracks, which run in series and conclude on Nov. 19.

The first track consists of five kernel problems. These cover where the model spends its time: linear projections, attention, the mixture-of-experts layers, and full decode and prefill layers. Contestants compete to write kernels faster than the SGLang baseline.

The second track is the serving track. Contestants directly modify SGLang. The kernel optimizations from the first track are a natural place to start. Runs serve the model on a full 8×H200 node, replaying multi-turn conversations at several load levels. Good submissions must push the throughput-interactivity frontier beyond that of SGLang.

The challenge runs on H200s. While much optimization work targets NVIDIA’s latest Blackwell GPUs, the previous generation, Hopper, is commonly used at universities and labs. Faster kernels here will make the best American open-weight model cheaper to serve and study for many users and researchers who are likely to build on it.

Submissions are licensed under the Apache License 2.0, SGLang’s license, so the best work can be iterated on and upstreamed into a serving system that people already use.

Compete

To get started, visit the challenge site for additional details and rules, as well as a download for the challenge CLI and instructions for interacting with the challenge API.

This challenge is made possible by Andromeda, which is providing compute. Andromeda makes compute accessible and reliable at any scale – across training and inference, short-term and long-term commitments, spot and reserved capacity, and from providers all over the world.

Thanks to Arjun Balaji, Ricardo de Arruda, and Alex Nails for helpful feedback.