Meta launches open-weight AI model designed to run on Mac computers

10:00

By: Azzat Manal

Meta launches open-weight AI model designed to run on Mac computers

Meta has introduced Muse Spark Open, a new open-weight artificial intelligence model designed to run locally on Apple Mac computers, expanding access to advanced AI tools for developers and professionals. The company describes it as the most capable US-developed model of its kind, highlighting its performance in complex coding tasks, multi-step reasoning and AI-driven workflows.

The model contains 816 billion parameters and supports a context window of up to one million tokens, according to Meta. It is natively multimodal, allowing it to process text and images, while its architecture is designed to support demanding software engineering tasks and workflows involving AI agents. These capabilities could help developers analyse large codebases, implement changes and automate parts of their work.

Running the model locally, however, requires substantial computing resources. According to reporting by Engadget, the system needs approximately one terabyte of unified memory, making a configuration involving two high-end Mac Studio computers equipped with Apple’s M5 Ultra chips particularly suitable. The hardware requirements place the model beyond the reach of most standard personal computers, despite its availability to developers.

Meta has also integrated Muse Spark Open with Muse Code, its AI-powered coding tool designed to assist with programming tasks. The company says it has tested the model across a range of agent-based workflows, including fixing software bugs, introducing new features and completing builds for a client application associated with the social platform Mastodon.

The model has also been used with DaVinci Resolve, professional video-editing software. In demonstrations described by Meta, the AI was able to review footage and crop video segments, suggesting potential applications beyond conventional programming. Such capabilities could make AI tools increasingly useful across creative and technical professions, although their practical value will depend on performance, reliability and the computing resources available to users.

Open-weight models offer a different approach from fully proprietary AI systems. By making trained model weights available, developers can download, adapt and integrate the technology into their own applications rather than relying exclusively on a provider’s hosted services. This can offer greater flexibility and control over data, although access to model weights does not necessarily mean that all training data, code or licensing terms are open.

Meta has made Muse Spark Open available through several application programming interface providers, while its model weights can be obtained through platforms including Hugging Face and major cloud services. The release gives developers different options for accessing the technology, whether through hosted services or by deploying it within their own computing environments.

The launch also highlights the growing competition among technology companies to make increasingly powerful AI systems accessible beyond large cloud data centres. By combining its model with Apple’s high-memory desktop hardware, Meta is promoting a setup capable of handling demanding workloads locally, while Apple gains an opportunity to demonstrate the computing capabilities of its latest chips.

Despite the potential benefits, the high memory requirements and specialised hardware needed to run Muse Spark Open could limit adoption among individuals and smaller organisations. Its longer-term impact will depend on whether developers can translate its capabilities into practical applications while managing the cost, complexity and energy demands of local AI deployment.