Meta's Muse Spark Open Coding Model Lands in Amazon SageMaker JumpStart

What AWS announced

AWS says Meta's Muse Spark Open, a coding-optimized open-weight model, is now available in Amazon SageMaker JumpStart, the machine learning hub within Amazon SageMaker AI. According to AWS, customers can deploy the model to a SageMaker AI endpoint in their own AWS account in a few steps, without managing the underlying serving infrastructure.

SageMaker JumpStart offers two variants, per AWS: a BF16 version called Muse-Spark-Open-816B-A42B, and an NVFP4 quantized version called Muse-Spark-Open-816B-A42B-NVFP4.

About the model

Meta describes Muse Spark Open as a coding-optimized model trained for agentic coding workflows, built with a focus on first-attempt accuracy, tool calling, and understanding of large codebases. AWS says it is a reasoning model that accepts text, images, documents, audio, and audio-video as input, and returns text as output.

The model uses a mixture-of-experts architecture with 816 billion total parameters, of which 42 billion are active for each token, according to AWS. It has a context window of 1 million tokens. AWS says Meta trained the model on real developer workflows such as multi-file refactors and long debugging sessions, so the model can take in an entire project, including dependencies and legacy code, without a task needing to be split into smaller pieces.

Meta says the model plans, acts, observes the result, and corrects itself over long horizons rather than making single tool calls. As a reasoning model, AWS says it works through a task before acting, and the SageMaker AI endpoint returns that reasoning alongside the answer. Meta trained the model to browse the web, coordinate subagents for tasks that can be split up, and operate computers across desktop, mobile, and browser environments, according to AWS, though on SageMaker AI the application or agent framework must supply the actual tools, browser, and computer-use environment. Meta also reports that the model reasons across images, documents, audio, and video.

Meta releases the weights under the Muse Model License, which includes the Muse Spark Usage Policy, AWS says.

Deployment details

AWS says the two variants require different instances because of memory needs: sparse activation keeps per-token compute low, but serving the model still requires enough GPU memory to hold all 816 billion parameters. The NVFP4 variant stores weights in a 4-bit floating point format that needs roughly a quarter of the memory of the BF16 weights. As a result, AWS says the NVFP4 variant runs on an ml.p5e.48xlarge instance, while the BF16 variant runs on an ml.p6-b300.48xlarge instance.

Because the weights are open and the endpoint runs in the customer's own AWS account, AWS says customers can attach it to their own Amazon Virtual Private Cloud, encrypt the storage volume with a key from AWS Key Management Service, and turn on network isolation for the container. Customers also choose the instance type and scaling limits and pin the model version, so AWS says the model only changes when the customer redeploys it.

AWS lists suggested use cases including coding assistants that work over private repositories, long-running coding agents that edit across multiple files and run tests, and multimodal engineering workflows that turn screenshots, design documents, or recorded bug reports into a reproduction plan or draft fix.

AWS says the model can be deployed from Amazon SageMaker Studio or with the SageMaker Python SDK, each creating a real-time SageMaker AI endpoint in the customer's account.

Source: AWS Machine Learning Blog, "Meta's Muse Spark Open now available in Amazon SageMaker JumpStart," published October 15, 2026.