Meta Releases Muse Spark Open AI Model for Local Deployment on Macs

Meta has released Muse Spark Open, an artificial intelligence model built for complex coding, visual analysis and multi-step tasks, giving developers access to its model weights for custom deployment. The release, announced on 15 October, includes support for running the model locally on a cluster of high-end Apple Mac Studio computers.

Developed by Meta Superintelligence Labs, Muse Spark Open has 816 billion parameters, with approximately 42 billion active for each token processed. It also supports a context window of up to one million tokens, allowing it to work with large amounts of information, including extensive codebases and lengthy documents.

Local AI Comes With Significant Hardware Requirements

Meta worked with Apple to optimise Muse Spark Open for Apple Silicon through the MLX machine-learning framework. The setup allows developers to run the model on their own hardware rather than relying exclusively on cloud-based AI services.

However, local deployment requires substantial computing resources. Apple's published setup uses two Mac Studio computers equipped with M5 Ultra chips and 512GB of unified memory each. An alternative configuration uses four Mac Studio systems with 256GB of unified memory each. The computers work together to distribute the model's processing workload.

This makes the release more relevant to AI developers, research teams and organisations with access to high-performance hardware than to ordinary MacBook owners.

Built for Coding and Video Workflows

Muse Spark Open is designed to handle tasks that require more than generating text or answering questions. Its capabilities include reading code, planning software changes, using tools and checking results as it works through a task.

Meta demonstrated the model working with Muse Code, its AI coding terminal, to make changes to Ice Cubes, an open-source Mastodon client. The workflow included fixing bugs, adding features and building the application.

The company also tested the model with DaVinci Resolve, the video-editing software. In that demonstration, the AI reviewed footage frame by frame to identify suitable moments and prepare vertical crops. These examples show how the model can assist with connected tasks inside software applications, although results will depend on the workflow and deployment setup.

The model can process text, images, video and audio inputs, giving developers options for building applications that work across different media formats. Its large context window is also intended to help with extended tasks that require the model to retain information across multiple steps.

Developers Can Download the Weights

Meta has made Muse Spark Open available through Hugging Face and several cloud and AI infrastructure providers, including Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry. Developers can also access hosted inference through providers such as Fireworks AI, Together AI, Baseten and Modal.

The release is open-weight rather than fully open-source software with unrestricted use. Developers can inspect and customise the released weights, but use is governed by Meta's Muse Model License and its acceptable-use requirements. Commercial users should check the licence terms before deploying the model in products or services.