Meta's Muse Spark Open lands in Amazon SageMaker JumpStart

Muse Spark Open stands out for how it was trained: Meta says it was built on real developer workflows such as multi-file refactors and long debugging sessions, and designed as an agentic model from the start, planning, acting, and correcting itself over long horizons. Its 1 million-token context window is meant to hold an entire project, dependencies and legacy code included, so a task need not be split to fit.

Serving it is a memory question. Sparse activation keeps per-token compute low, but the endpoint still has to hold all 816 billion parameters, so instance choice comes down to capacity. That is why the NVFP4 variant, which stores weights in NVIDIA's 4-bit floating point format, becomes the one to try first when sizing for cost, even though the walkthrough advises benchmarking both variants on your own repositories before picking a production instance.

Control is the other half of the pitch. Because the endpoint runs in the customer's account, they can attach it to their own Amazon VPC, encrypt storage with an AWS KMS key, turn on network isolation, and pin the model version so it only changes on redeploy. AWS notes that deployment typically takes 15–20 minutes, and that hosted instances are billed for as long as they run, so the post walks through deleting the endpoint when finished.