Meta has released Muse Spark Open, a new open-weight artificial intelligence model designed for advanced coding, complex software engineering and autonomous AI workflows. Announced on October 15, 2026, the model allows developers to download its weights, customize it and run it on supported computing infrastructure.
According to Meta, Muse Spark Open is built to handle lengthy, multi-step tasks that require planning, tool use and repeated checks. The release expands the company’s efforts to give developers greater control over how they deploy and customize AI models.
What is Meta Muse Spark Open?
Muse Spark Open is an 816-billion-parameter model built using a sparse mixture-of-experts architecture. It activates approximately 42 billion parameters per token, allowing it to use a subset of its overall model capacity for each processing step.
Related reading: Meta’s New Muse AI Models: What You Need to Know
The model supports text, image, video and audio inputs, alongside a context window of approximately one million tokens. This enables it to process large amounts of information within a single context, including extensive codebases and supporting documentation.
Meta says the model is designed for agentic work, where AI systems can plan tasks, use tools, inspect results and make corrections instead of simply responding to individual prompts.
How Muse Spark Open helps developers
Muse Spark Open is designed to support complex software engineering tasks, including navigating codebases, implementing changes, fixing bugs and validating results.
Developers can use it to build coding agents that break larger objectives into smaller steps, interact with development tools and revise their approach when something goes wrong. Meta has also demonstrated workflows involving changes to an iOS application and video editing in DaVinci Resolve.
These capabilities could make the model useful for developers seeking more control over automated coding workflows. However, its real-world performance will depend on the task, deployment environment and independent evaluations.
Running the model on desktop hardware
One notable aspect of the release is its optimization for Apple’s Mac Studio computers equipped with M5 Ultra chips.
Meta says two such systems can run the model using Apple’s MLX framework and their combined unified memory. This setup is intended to bring advanced AI inference to local desktop hardware rather than relying entirely on cloud-based processing.
The hardware requirements remain substantial, meaning this configuration is aimed primarily at developers and organizations with access to high-end computing equipment.
Where developers can access Muse Spark Open
Meta has made the model weights available through platforms including Hugging Face, Amazon SageMaker AI.
