Meta and Apple enable local run of 816B multimodal model on dual Mac Studios
Meta and Apple partner to enable local execution of Meta’s largest open-weight model on two M5 Ultra Mac Studios using the MLX framework.
Key points
- Meta released Muse Spark Open, an 816B-parameter multimodal model with 42B active parameters and a 1M-token context window.
- The model is optimized to run locally on two M5 Ultra Mac Studios linked via Apple’s unified memory architecture and MLX framework.
- Meta claims this is the most capable US open-weight model, designed for complex coding and agentic tasks.
Meta released the weights for Muse Spark Open today, describing the system as the most capable US open-weight model available. The company collaborated with Apple to optimize the large multimodal model for local execution on Apple Silicon hardware, specifically targeting the Mac Studio desktop line. This release marks a shift toward high-performance local AI deployment for developers and researchers.
The model contains 816 billion total parameters, with 42 billion active parameters during inference. It features a one-million-token context window and is designed for coding, software engineering, and complex agentic work. According to Meta, the system excels at long-horizon tasks and multi-step reasoning. The weights are available for download on Hugging Face and other platforms.
View the original post on X
Hardware and Software Integration
Meta states that Muse Spark Open is optimized to run on the Mac Studio powered by the M5 Ultra chip. The setup requires two Mac Studios to host the entire model. Apple’s unified memory architecture allows each machine’s integrated GPU to access up to 512 GB of memory. Combined, the two machines provide a terabyte of unified memory.
The collaboration leverages Apple’s latest version of MLX, an open-source array framework. MLX is optimized for the memory bandwidth of Apple silicon chips. Meta explains that the M5 Ultra’s 1.2 TB/s memory bandwidth is sufficient to read the small data slice needed for each token. This configuration makes the two desktops behave like a single supercomputer.
Industry Context and Availability
The partnership between Meta and Apple is notable given their historical competitive tensions. Meta AI chief Alexandr Wang commented on the cooperation in an interview with Axios. He expressed excitement about the companies collaborating, using a colloquial phrase about them playing nice again. Wang explained that Apple and Meta have been working together for several months.
Muse Spark Open is free to download and includes a license allowing limited commercial use. The model is accessible via API platforms including Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI. Weights are also available through Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models. Apple has published an article on its Machine Learning Research website highlighting local experimentation capabilities.
What should you do?
Security and technology professionals should monitor the availability of the 512 GB Mac Studio models, as Apple has not yet released this specific configuration. Developers interested in local large language model deployment should review the MLX framework documentation. Organizations using agentic AI workflows should assess whether local processing meets their latency and privacy requirements compared to cloud alternatives.
Background: Large language models
Large language models are a specific type of neural network trained on text. Generative AI is the broader umbrella term for any system that creates new content, including images and code. Not all generative AI uses large language models, but most modern text generators do.
What to do and how to stay safe: Meta
- Verify the system requirements for running 816B-parameter models locally, ensuring sufficient unified memory and bandwidth.
- Review the license terms for Muse Spark Open to confirm compliance with limited commercial use restrictions.
- Monitor Apple’s hardware release schedule for the 512 GB Mac Studio configuration required for full model deployment.
- Test local inference latency using the MLX framework to determine if on-device processing meets operational needs.
Step-by-step guide: Mobile Device Security: How to Protect Data Beyond Passwords
General security guidance from the Threat Times newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
How many Mac Studios are needed to run Muse Spark Open locally?
Two Mac Studios with M5 Ultra chips and 512 GB of memory each are required to hold the entire model.
Is Muse Spark Open available for commercial use?
Yes, the model is free to download and has a license that allows limited commercial use.
What framework does Meta use to optimize the model for Apple Silicon?
Meta uses Apple’s MLX framework, which is optimized for the unified memory architecture of Apple silicon.
