Meta Launches Muse Spark Open, a Powerful AI Model for Apple Macs

Meta has released a new, highly capable open-weight AI model designed to run locally on Apple silicon Macs. This collaboration between Meta and Apple, reported by TechCrunch, marks a significant step in bringing frontier-model performance to personal workstations, particularly for developers and creative professionals.

What is Muse Spark Open?

Meta's new model, called Muse Spark Open, is an open-weight, multimodal AI system. According to the company's announcement, it leverages 816 billion parameters and supports a context window of 1 million tokens. Meta claims its evaluation results show superior performance compared to other popular open-weight models. The model is natively multimodal and is built with a focus on complex coding tasks and "agentic labor," meaning it can autonomously execute multi-step workflows.

Why is it built for Macs?

In a notable industry alignment, Meta designed Muse Spark Open to run locally on Apple computers. This optimization for Apple silicon suggests deep technical collaboration, allowing the model to tap directly into the unified memory architecture of Macs. However, according to a report by Engadget cited in the TechCrunch article, running the model locally demands substantial hardware: at least 1 terabyte of RAM. Meta indicates it will run best on a Mac Studio equipped with the M5 Ultra chip, positioning it as a tool for high-end professional setups rather than consumer laptops.

What can it actually do?

The source material provides concrete examples of the model's applied capabilities. Meta paired Muse Spark Open with its own Muse Code AI coding terminal to deploy it in agentic workflows. In one instance, it was used to implement coding fixes for a Mastodon client, handling "bug fixes, new features, [and] full build." Perhaps more surprisingly, the company also deployed it to run DaVinci Resolve, the professional video editing software, where the model demonstrated the ability to review and crop footage autonomously. This points to a vision of AI as a direct actor within complex creative and development software.

What are the implications of an "open-weight" model?

Open-weight models, where the model's parameters (or "weights") are publicly released, offer distinct advantages over closed, proprietary systems like those from OpenAI or Google. As highlighted in the reporting, they can be more cost-efficient for large-scale deployment and offer greater transparency and flexibility. Developers can audit, modify, fine-tune, and run them without vendor-locked API costs. This release continues Meta's strategy of advancing open AI research, though the hardware requirements for local execution create a high barrier to entry for most users.

How does this relate to the AI tools in Grafy?

While Grafy does not run local models of this scale, the themes of open models, agentic workflows, and multimodal AI are central to the platform. For developers and creators interested in exploring the capabilities hinted at by Muse Spark Open, several Grafy apps provide accessible, cloud-based avenues.

For coding and agentic tasks, Grafy Code is an AI coding studio where you can experiment with various cloud models to generate, debug, and run code in a sandboxed environment. To test multimodal reasoning on documents or research papers, you can use Reader to upload PDFs and interrogate them with AI. Furthermore, the Graph app allows you to build, visualize, and iterate on complex AI-powered workflows, mirroring the "agentic" pipelines Meta described.

For a deeper look at the context of this Apple-Meta collaboration, you can read our previous analysis: Meta and Apple collaborate on Muse AI models for Apple Silicon Macs. The release also connects to broader discussions about AI efficiency and local execution, a topic explored in our coverage of the Janas LLM engine.

Frequently asked questions

What hardware is needed to run Muse Spark Open locally?

According to the Engadget report cited by TechCrunch, running Muse Spark Open locally requires at least 1 terabyte of RAM. Meta states it runs best on a Mac Studio with an M5 Ultra chip, making it a tool for high-end professional workstations, not standard consumer hardware.

How is this different from using an AI API?

Using an open-weight model locally, like Muse Spark Open on a capable Mac, eliminates ongoing API costs and latency, offers full data privacy, and allows for model modification. However, it requires a massive upfront hardware investment versus the pay-as-you-go, no-setup nature of cloud APIs.

What does "agentic labor" mean for this model?

As per Meta's examples, "agentic labor" means the model can autonomously execute multi-step tasks. This includes fixing bugs and adding features to a codebase or operating video editing software like DaVinci Resolve to perform specific edits without step-by-step human guidance.

---

Source: TechCrunch — AI — Meta launches an open-weight AI model that runs on Macs