Apple and Meta have announced a collaboration that allows developers to run one of Meta’s largest AI models locally on Mac computers, reducing the need for cloud-based processing.
At the center of the partnership is Muse Spark Open, an 816-billion-parameter open-weight model designed for coding, video analysis and AI agents. Supporting workloads of this size requires several high-end Mac Studio computers connected to share processing demands.
The partnership is an unexpected development for two longtime competitors whose disagreements over privacy and advertising have strained their relationship for years.
Why It Matters: Running advanced AI models on desktop hardware gives organizations more control over where their data is processed and how AI workloads are managed. For enterprises investing heavily in AI, the ability to handle demanding workloads internally could reduce recurring cloud expenses and provide greater flexibility in meeting security and governance requirements.
- Meta Releases an 816-Billion-Parameter Open-Weight Model: Muse Spark Open supports a context window of up to 1 million tokens, allowing it to handle extensive codebases and information-heavy tasks. Developers can download the model’s weights and deploy them on their own infrastructure under a license permitting limited commercial use. Meta’s proprietary Muse Wonder remains more advanced, although the open-weight release gives developers greater flexibility to evaluate and customize the technology.
- Local Deployment Requires a Significant Hardware Investment: Supporting Muse Spark Open requires between two and four Mac Studio computers equipped with Apple’s M5 Ultra processors and approximately 1 terabyte of combined memory. A Mac Studio with 256GB of memory starts at $10,799, putting a four-machine configuration above $43,000 before additional expenses. Apple’s MLX framework distributes processing across connected computers, allowing them to handle models too large for an individual machine. The cost makes this setup more suitable for specialized development environments than widespread deployment.
- Cloud Economics Introduce Another Consideration: Local inference could reduce recurring cloud expenses for organizations with consistent AI workloads, although the upfront investment makes utilization an important factor. Dedicated hardware may be harder to justify for occasional development projects, where cloud capacity offers greater flexibility. Data residency and control over sensitive workloads could also influence deployment decisions, even when local processing is not the least expensive option.
- Years of Competition Give Way to Cooperation: Apple and Meta have had a strained relationship for years, particularly after Apple’s 2021 privacy changes restricted advertising tracking. Meta estimated the changes would cost the company approximately $10 billion in revenue in 2022. Despite their history, engineering teams spent months working together to optimize Muse Spark Open for Apple silicon. Neither company has confirmed whether the collaboration will extend to other products.
- Early Testing Explores Agentic Workflows: Meta demonstrated Muse Spark Open completing software development tasks, including fixing bugs, adding features and building an application. In another test, the model reviewed and cropped video footage using editing software. These demonstrations show how locally deployed AI agents could work within existing applications, although their performance in enterprise production environments remains untested.
Go Deeper -> Apple and Meta team up to make Muse AI model run on Mac computers – CNBC
Meta launches an open-weight AI model that runs on Macs – TechCrunch
