Meta releases 816B Muse Spark Open model for local Apple Silicon deployment
Meta publishes weights for its largest open-weight AI model, enabling multimodal agentic tasks on dual Mac Studios with 512GB memory each.
Key points
- Muse Spark Open features 816 billion total parameters with 42 billion active and a one-million-token context window.
- The model runs locally on two Apple M5 Ultra Mac Studios using the MLX framework and unified memory architecture.
- Weights are available on Hugging Face, Amazon SageMaker, Google Cloud, and Microsoft Foundry, with API access via seven providers.
Meta released the weights for Muse Spark Open today, marking the debut of its most capable US open-weight model. According to the company’s announcement on Hacker News, the release aims to provide advanced agentic capabilities directly on desktop hardware, specifically targeting Apple Silicon users who wish to run frontier models locally without relying on cloud infrastructure.
The model contains 816 billion total parameters, with 42 billion active parameters during inference. It supports a context window of one million tokens and is designed for multimodal tasks. Meta states that Muse Spark Open excels at long-horizon tasks, complex software engineering, and multi-step reasoning, positioning it as a tool for developers requiring sustained logical processing and code generation capabilities on their own machines.
Technical Specifications and Hardware Requirements
Running this model locally requires significant hardware resources. Meta optimized Muse Spark Open to operate on two Apple M5 Ultra Mac Studios simultaneously. According to the source material, Apple’s unified memory architecture allows each M5 Ultra chip to utilize up to 512 gigabytes of memory. By connecting two machines, users can create a system with a terabyte of unified memory, effectively behaving like a single supercomputer.
The M5 Ultra chip provides 1.2 terabytes per second of memory bandwidth, which Meta says is sufficient to read the small data slice required for each token. The company leveraged the latest version of the MLX framework to enable this performance. This setup allows the integrated GPU to manage the heavy computational load while keeping data processing entirely on-device, avoiding network latency associated with cloud-based inference.
Availability and Developer Ecosystem
Developers can access Muse Spark Open through multiple channels today. API platforms including Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI offer immediate access. For those wishing to host the model themselves, weights are available for download on Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face, and Microsoft Foundry Models.
The model supports leading inference engines such as SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, Apple MLX, llama.cpp, and Unsloth. Meta notes that optimizations are included for AMD, Apple, Intel, and NVIDIA silicon. Developers can further customize the model using the PyTorch framework, allowing for tailored implementations in specific enterprise or personal workflows.
Demonstrated Use Cases
Meta demonstrated the model’s capabilities through on-device agentic workflows using Muse Code, which is powered by Muse Spark Open. The company reported shipping real iOS changes to Ice Cubes, an open-source Mastodon client. These changes included bug fixes, new features, and full build processes, all handled by the model running locally.
Another test involved driving DaVinci Resolve end-to-end. The model reviewed video footage frame by frame to identify the strongest moments and cropped them vertically. According to the source, these tests highlight the model’s ability to handle complex, multi-step creative and engineering tasks without external cloud assistance, reinforcing Meta’s stated goal of making advanced models accessible to everyone.
What should you do?
Security and technology professionals should evaluate whether local deployment of large language models aligns with their data privacy requirements. Organizations handling sensitive code or proprietary data may benefit from the air-gapped nature of on-premise AI inference. However, teams must verify that their hardware infrastructure can support the memory and bandwidth demands of models with hundreds of billions of parameters before attempting local implementation.
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
- Assess if your organization has the hardware capacity, specifically high-memory Apple Silicon or equivalent, to run large local models.
- Review the security implications of downloading and hosting model weights from public repositories like Hugging Face.
- Test the model’s performance on your specific agentic workflows using the provided inference engines before full deployment.
- Monitor the evaluation methodology report for details on the model’s capabilities and limitations in reasoning tasks.
Step-by-step guide: Prompt Injection Warning Signs: How to Spot AI Manipulation
General security guidance from the Threat Times newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
What is the size of the Muse Spark Open model?
The model has 816 billion total parameters with 42 billion active parameters.
What hardware is required to run Muse Spark Open locally?
It requires two Apple M5 Ultra Mac Studios with up to 512 GB of memory each.
Where can I download the model weights?
Weights are available on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models.
