Meta’s Muse Spark Open Brings a Huge AI Model to Local Macs

Meta has released Muse Spark Open, an open-weight model the company says is built for complex coding and agentic tasks and can run locally on Apple computers. The model has 816 billion parameters and a one-million-token context window, according to TechCrunch’s report. Meta also published evaluation results comparing the model with other systems; those are company-reported benchmarks, not a universal guarantee of performance.

The word “local” needs context. The report says running Muse Spark Open requires at least one terabyte of RAM and that it is expected to work best on a high-end Mac Studio configuration. That places the model outside the reach of most ordinary laptops and desktops. The launch is significant for experimentation with local inference, but it does not mean a typical Mac owner can download the weights and run them comfortably.

Why open weights matter

Open weights let researchers and developers inspect and run a model within certain license terms, adapt it for their needs, and potentially host it on their own infrastructure. That can provide more control over data flow and deployment than a hosted chatbot. It can also shift costs to hardware, power, setup, and ongoing maintenance. “Open-weight” should not automatically be read as fully open source: training data, development process, and other components may not all be public.

Meta says Muse Spark Open is multimodal and designed for coding and agentic workflows. It has paired the model with its Muse Code terminal and described using it for coding tasks. Developers should still test the model against their own workloads, review license terms, and compare results on accuracy, latency, and total cost. A large parameter count or long context window alone does not establish that a model is best for a particular job.

What to watch next

The practical test will be whether usable software and hardware configurations make the model accessible beyond specialized labs. Community tooling, memory optimization, and independent evaluations will shape that answer over time. Until then, local deployment is a powerful option for well-equipped teams, not a frictionless replacement for cloud AI.

See Talluqat’s Technology news and our article about another open-weight model release.

Reporting basis: Meta’s launch and specifications as reported by TechCrunch; performance claims are attributed to Meta.