Meta Releases Muse Spark Open: AI Coding-Optimized Model Now Available on SageMaker JumpStart
Summary:Meta has officially released Muse Spark Open, a coding-optimized open-weight model, now available on Amazon SageMaker JumpStart. The model features a Mixture of Experts (MoE) architecture with 816 billion total parameters (42 billion active per token) and supports a 1 million-token context window. It can process multimodal data, including text, images, documents, audio, and audio-video, and is designed for agentic workflows, focusing on first-attempt accuracy, tool calling, and understanding lar
Meta Releases Muse Spark Open: AI Coding-Optimized Model on SageMaker JumpStart
Meta has announced the release of Muse Spark Open, an open-weight model optimized for coding and agentic workflows, now available on Amazon SageMaker JumpStart. This release represents a significant advancement in AI-assisted programming and intelligent agent development.
Architecture and Key Features
Muse Spark Open employs a Mixture of Experts (MoE) architecture with a total of 816 billion parameters, but only 42 billion are active per token. This design maintains the capacity of a large network while significantly reducing the computational overhead for each token. The model supports a 1 million-token context window, allowing it to process entire projects, including dependencies, legacy code, and thousands of files, without the need to split tasks into smaller pieces.
Additionally, Muse Spark Open is designed as a reasoning model, accepting text, images, documents, audio, and audio-video as input and returning text output. Its applications include:
- Coding Assistants over Private Repositories: Answering questions about large codebases, refactoring, and debugging code processed on an endpoint in the user's account.
- Long-Running Coding Agents: Running agents that plan changes, edit across multiple files, run tests in the user's sandbox, and adjust based on the results.
- Multimodal Engineering Workflows: Transforming screenshots, design documents, or recorded bug reports (including audio and video) into reproduction plans or draft fixes.
Deployment and Usage
Muse Spark Open offers two variants in SageMaker JumpStart: BF16 and NVFP4 quantized versions. The BF16 version runs on the ml.p6-b300.48xlarge instance, while the NVFP4 version runs on the ml.p5e.48xlarge instance. The NVFP4 variant, utilizing NVIDIA's 4-bit floating-point format, requires roughly a quarter of the memory of the BF16 variant, making it the preferred choice for cost-sensitive applications.
Users can deploy the model using SageMaker Studio or the SageMaker Python SDK. The deployment process involves selecting the model, accepting the license agreement, configuring the endpoint instance type and count, and choosing advanced options such as VPC and AWS KMS key. Once deployed, users can send inference requests, and the model returns the inference result along with an explanation of the reasoning process.
Developer Recommendations
When choosing the production instance, it is recommended to benchmark both variants to ensure the quality meets your use case. Additionally, due to the large parameter size of the model, ensure that the GPU memory capacity of the instance is sufficient.
Conclusion
The release of Muse Spark Open provides developers with a powerful AI coding assistant, particularly suited for handling complex codebases and long-running tasks. Through SageMaker JumpStart, developers can easily deploy and use the model, and integrate it with other AWS services to build smarter applications.
Tags
- Meta
- Muse Spark Open
- Mixture of Experts
- Coding Assistant
- Multimodal AI
Reading Time
8 minutes
— END —Source: AWS Machine Learning Blog (2026-10-15)
Tags: #Meta #Muse Spark Open #Mixture of Experts #Coding Assistant #Multimodal AI
Editorial & Fact-Checking Note: This article is compiled from primary research, official release documentation, and source papers by the ZICQ Newsroom pipeline with automated entity verification and human editorial review. If you notice any technical inaccuracy, please submit a correction via our corrections policy or email our editorial desk directly.
Community Comments