"Mommy and daddy" play nice: Apple and Meta team up to bring open-source Muse to Macs

Meta is releasing Muse Spark Open, an open-weight AI model it calls the most powerful U.S.-developed model of its kind. Apple and Meta worked together to make it run locally on Apple's highest-end Macs. The deal matters because it creates an unexpected partnership between longtime rivals and shows how AI is reshaping technology alliances.

The model runs across two or four connected Mac Studios powered by Apple's M5 Ultra chips. It uses Apple's MLX framework, which helps the model operate on Apple hardware instead of relying only on remote cloud computers. Meta will offer the model free to download, with a license allowing commercial use.

This is currently a demonstration and proof of concept, not an announced Apple Intelligence integration. Meta benefits from wider model adoption, while Apple can highlight its computers' capabilities. The companies still compete, but they say collaboration makes sense where their interests overlap.

An open-weight AI model makes its trained weights available for others to download, run, adapt, or inspect. The weights are the numerical settings produced during training. This gives companies more control than using a closed service, where the provider controls access and operation. The article describes open-weight systems as a way to cut costs and offer alternatives to proprietary models.

A fully open-source model generally goes further. It may provide the software code, training methods, and sometimes training data or detailed development information. Open-weight release alone does not promise that all those elements are available. Muse Spark Open is described as open-weight, and Meta plans a license allowing commercial use.

A proprietary model keeps important components controlled by its owner. Users may access it through a product or cloud provider but cannot freely obtain or modify the model itself. The article does not define these categories in detail, so this distinction uses standard AI terminology.

Muse Spark Open has 816 billion parameters, making it an exceptionally large AI model. Parameters are the internal values a model adjusts during training to represent patterns. More parameters can support greater capability, though size alone does not determine how well a model performs. Meta presents this system as the most powerful U.S.-developed open-weight model available.

Apple and Meta demonstrated the model running locally across two or four connected Mac Studios. The computers are powered by Apple's M5 Ultra chips and use Apple's MLX framework. Connecting multiple machines provides the computing resources needed to handle a model of this scale without sending each prompt to a remote service.

The demonstration is important because it tests whether powerful AI can run on personal or workstation hardware. It does not mean every Mac can run the model. Instead, it shows what high-end Apple computers can do and hints at more local AI systems built by companies.

Running an AI model locally means prompts and processing stay on the user's own computers instead of being sent to a model provider's servers. That matters for privacy. In this collaboration, Meta says it has no control or visibility into how developers use Muse Spark Open when it runs locally. This directly sidesteps a long-running Apple-Meta disagreement over data and privacy.

The example is the model operating across connected Mac Studios with M5 Ultra chips. Developers can download it, run it on their own hardware, and use a license permitting commercial applications. They do not need to send each request to Meta's cloud. Keeping operation local can also give organizations more control over deployment and data handling.

Local execution is not automatically risk-free. Open access can let bad actors remove safeguards and pursue cyberattacks or biological threats. Still, the partnership shows how local AI can combine privacy, developer control, and powerful capabilities, while Apple continues deciding whether to add Meta models to Apple Intelligence.

Powerful open-weight models can change AI economics by giving companies an alternative to buying access to proprietary systems. A downloadable model can be adapted and operated under the user's control. The article says these systems promise to slash AI costs, provide more control, and offer an alternative to the Chinese AI ecosystem. Commercial licensing can make that option useful to businesses.

Muse Spark Open provides a concrete example. Its 816 billion parameters run locally across two or four M5 Ultra-powered Mac Studios. That proof of concept suggests some organizations could build their own AI systems using different hardware and deployment choices, rather than depending entirely on Nvidia GPUs or leading AI labs' proprietary models.

The shift remains a possibility, not a settled result. Large models still require substantial computing resources, and the article does not quantify savings. Meta is also making the model available through cloud providers and developer tools. Wider adoption will show whether open-weight systems genuinely alter costs and competitive power.

Apple and Meta still compete over devices, privacy, advertising, and the future of computing. Yet AI creates a shared opportunity. Meta wants its models widely available, and Apple wants to demonstrate what its most powerful computers can do. Working together on a technical showcase serves both goals without ending their broader rivalry.

Daniel Gross, a former Apple executive now at Meta, helped bring the companies together. Their project makes Muse Spark Open run locally across connected Mac Studios. Because the model operates on the computers, Meta does not receive users' prompts or other information. That helps avoid the privacy issue that has long divided the companies.

The détente is limited. Apple and Meta have not resolved their disagreements, and they continue competing in several markets. A bigger future step would be Apple incorporating Meta models into Apple Intelligence. For now, the partnership shows that rival companies can collaborate selectively when AI makes their interests overlap.

Parameters are adjustable numerical values inside an AI model. During training, the model changes them as it processes examples. Together, they encode patterns about how information relates. When a user asks a question, the model uses those learned settings to estimate and generate a useful sequence of answers. The article identifies Muse Spark Open's size but does not explain parameters directly.

Muse Spark Open has 816 billion parameters. That enormous number helps explain why Meta presents it as a highly capable model and why it needs two or four connected Mac Studios with M5 Ultra chips for local operation. The parameters are not separate facts or stored answers. They are settings that influence how the model responds.

A larger parameter count can provide more capacity to represent complicated patterns, but it does not guarantee better results by itself. Training quality, data, design, and safeguards also matter. In this case, the article's main point is that a model of this scale can now be demonstrated on Apple hardware rather than only remote systems.