What is Muse Spark Open?
Meta's coding-focused open-weight model Muse Spark Open is now accessible through Amazon SageMaker JumpStart. The model was trained on real developer workflows, such as multi-file refactors and long debugging sessions.
Key features
- 816 billion total parameters with 42 billion active per token (mixture-of-experts architecture)
- 1 million token context window: an entire project, its dependencies and thousands of files can be processed at once
- Text, image, document, audio and audio-video input; text output
- Agentic behavior: planning, acting, observing results and self-correcting
- Capabilities for web browsing, coordinating subagents and operating across desktop, mobile and browser environments
Two variants in SageMaker JumpStart
JumpStart offers BF16 (Muse-Spark-Open-816B-A42B) and NVFP4 quantized (Muse-Spark-Open-816B-A42B-NVFP4) versions. NVFP4 stores weights in NVIDIA's 4-bit floating point format, requiring roughly a quarter of the memory of BF16 weights. As a result, the NVFP4 variant runs on ml.p5e.48xlarge, while the BF16 variant runs on ml.p6-b300.48xlarge.
Why it matters
The model is released under the Muse Model License, which includes the Muse Spark Usage Policy. Because the SageMaker AI endpoint runs in your own AWS account, your source code and prompts are processed on your own infrastructure. Meta says the model also reasons across images, documents, audio and video beyond code.
