News in short

• Meta has introduced Muse Spark Open, an open-weight AI model designed to run locally on Macs.
• The model uses 816 billion parameters and supports a context window of up to one million tokens.
• Meta says it supports multimodal tasks, complex coding and agent-based workflows.
• Running it locally requires at least 1TB of RAM, making compatible hardware a major constraint.
• Meta says it tested the model for coding fixes and video-editing tasks, and has made it available through API platforms.

Meta has launched Muse Spark Open, an open-weight AI model designed to run locally on Apple’s Mac computers. The model packs 816 billion parameters and supports a context window of one million tokens.

Meta describes it as one of the most capable open-weight models developed in the United States. Its release highlights the company’s push to bring advanced AI capabilities to local computing environments.

However, running the model on a Mac will require substantial hardware resources. According to Engadget, users need at least 1TB of RAM to run it locally.

That requirement puts Muse Spark Open beyond the reach of most standard Mac configurations. Meta’s model is expected to work best on a Mac Studio equipped with the M5 Ultra chip, according to the report.

Muse Spark Open targets coding and AI agents

Muse Spark Open is natively multimodal, meaning it can work across different types of information. Meta says the model is built for complex coding tasks and agentic workflows.

AI agents can perform multi-step tasks with varying levels of autonomy. These workflows can involve analysing information, writing code, testing changes and completing tasks across software tools.

Meta says it paired Muse Spark Open with Muse Code, its AI coding terminal, to test these capabilities. The company used the model in several agent-based workflows, including work on a Mastodon client.

According to Meta, the model handled bug fixes, new features and a full build. These examples indicate how the company sees the model supporting more involved software development tasks.

Meta also tested Muse Spark Open with DaVinci Resolve, the professional video-editing application. The company said the model could review and crop footage.

These demonstrations point to potential uses beyond code generation. However, Meta’s reported results do not establish how the model performs across every application or real-world workflow.

Why running an open-weight model locally matters

Open-weight AI models give developers access to a model’s learned parameters. This can offer greater flexibility than relying entirely on a closed, hosted AI service.

Developers can integrate such models into their own tools and workflows. Depending on the licence and available resources, they may also adapt or deploy them in different environments.

Local execution can offer greater control over where data is processed. It may also reduce dependence on external inference services, although actual privacy and cost benefits depend on implementation.

Muse Spark Open’s hardware requirements present a significant trade-off. While local execution may appeal to developers and organisations handling complex workloads, 1TB of RAM is far beyond the memory available in most consumer computers.

The requirement also means that running the model locally could involve substantial hardware costs. Users without suitable machines may need to access the model through supported cloud or API services instead.

Meta makes the model available through APIs

Meta has made Muse Spark Open available through several API platforms. This gives developers another way to explore the model without necessarily running it on their own hardware.

The release adds to the growing range of open-weight models targeting coding, multimodal processing and AI agents. It also reflects a broader industry effort to make advanced models more adaptable to specialised workflows.

For now, Muse Spark Open’s local Mac support comes with a major qualification: the hardware needed to run it. Its practical appeal will depend on how its performance compares with alternatives and whether developers can justify the resources required.

Meta’s launch therefore offers two distinct propositions. Developers can experiment with the model through APIs, while users seeking local execution will need access to an unusually powerful Mac setup.