AWS adds Meta’s Muse Spark Open to SageMaker JumpStart

AWS has made Meta’s Muse Spark Open available through Amazon SageMaker JumpStart, giving customers a managed route to deploy the open-weight coding model to real-time SageMaker AI endpoints in their own AWS accounts.

Customers can choose between two variants with different precision and memory requirements. JumpStart offers the BF16 model as Muse-Spark-Open-816B-A42B on an ml.p6-b300.48xlarge instance. The NVFP4 model, Muse-Spark-Open-816B-A42B-NVFP4, runs on an ml.p5e.48xlarge instance.

AWS says customers can launch either variant from SageMaker Studio or through the SageMaker Python SDK. JumpStart handles the serving workflow, while customers retain control of the endpoint configuration in their accounts. According to AWS, they can connect the endpoint to a virtual private cloud, encrypt its storage volume with an AWS Key Management Service key, enable container network isolation, set scaling limits and pin a model version.

The lower-precision NVFP4 checkpoint has a smaller weight-storage footprint. Meta documents the BF16 checkpoint at approximately 1.65 TB and the NVFP4 checkpoint at approximately 537 GB. The figures exclude the key-value cache, activations and other runtime overhead, so they are not estimates of total deployment memory.

At launch, availability is limited to US East (N. Virginia) and US East (Ohio). AWS directs customers to check JumpStart for later regional availability but gives no timetable for expansion.

Muse Spark Open is a gated model, and AWS says users must accept its end-user license agreement before deployment. Meta says the model weights are released under the Muse Model License and an acceptable-use policy.

Meta describes Muse Spark Open as a sparse mixture-of-experts model with 816 billion total parameters and about 42 billion active parameters per token. In this architecture, each token is routed through a subset of the model instead of activating every parameter. Meta says the model has a 1,048,576-token context window, accepts text, images, audio and video, and generates text.

Meta says it trained Muse Spark Open for coding and long-running agent tasks, including reading large codebases, planning changes, calling tools, checking results and recovering from failed steps. These are Meta’s capability claims; DataPhoenix has not independently benchmarked them.

AWS warns that the multi-GPU SageMaker endpoint continues to accrue charges whenever it is running, including when it is not serving requests. The announcement links to SageMaker pricing but does not estimate deployment costs for either required instance type.