Meta’s Muse Spark Open now available in Amazon SageMaker JumpStart
Amazon Web Services Aris Tsakpinis ● Covered by 18 sources
Meta’s Muse Spark Open just landed in Amazon SageMaker JumpStart. AWS says you can run the coding model in your own account, with a 1M-token context window.
Based on reporting by Amazon Web Services, Aris Tsakpinis — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
AWS has added Meta’s Muse Spark Open to Amazon SageMaker JumpStart, giving customers a way to deploy the coding-focused open-weight model inside their own AWS accounts. The pitch is simple: use the model without running the serving stack yourself, then send it requests from SageMaker Studio or the SageMaker Python SDK.
Muse Spark Open was trained on real developer work, including multi-file refactors and long debugging sessions. Meta positions it as a reasoning model for agentic coding, with an emphasis on first-attempt accuracy, tool calling, and understanding large codebases. It takes text, images, documents, audio and audio-video as input, and returns text. Under the hood, it uses a mixture-of-experts setup with 816 billion total parameters and 42 billion active for each token.
The headline feature is the 1 million-token context window. That lets the model take in an entire project, plus dependencies, legacy code and thousands of files, without forcing users to chop the task into pieces. Meta also says the model can plan, act, observe results and correct itself over long horizons, and that it was trained to browse the web, coordinate subagents and operate computers across desktop, mobile and browser environments. On SageMaker AI, the surrounding application still has to provide those tools and environments.
AWS is offering two variants in JumpStart: BF16 and NVFP4. The smaller-memory NVFP4 version runs on ml.p5e.48xlarge, while the BF16 version runs on ml.p6-b300.48xlarge. AWS says NVFP4 is the one to try first if cost matters, though it also recommends testing both on your own workloads before betting production on either one.
The deployment story leans hard on control. The endpoint lives in the customer’s AWS account, can be attached to a VPC, can use KMS encryption, and can run with network isolation. AWS also lets customers pin a model version, which is a polite way of saying the model won’t surprise anyone by changing underfoot.
My take — AI-written commentary, not fact-checked reporting
This is the sane part of the open-model story: put the weights where the customer controls the endpoint, not in a mystery box with a cheerful marketing page. The catch is still the catch — giant models are only cheap in slide decks, and someone has to pay for those multi-GPU instances while they sit there doing absolutely nothing. Still, AWS is treating open weights like infrastructure, which is the right attitude for once.
Read more about this at: Amazon Web Services
