Meta's open-weight Muse Spark Open aims to run on Apple Macs
Meta says its new open-weight model, Muse Spark Open, can run locally on Apple silicon, though reports say it needs about 1TB of memory.
TechCrunch reports that Meta has launched a new open-weight AI model called Muse Spark Open, developed with support from Apple. According to the outlet, the model is designed to run locally on Apple silicon rather than only in a remote data center.
Engadget, in a separate report published earlier the same day, describes the model as a very large system and notes a significant hardware requirement. The outlet reports that running it would take roughly 1TB of RAM, a memory footprint far beyond what typical Mac owners have in their machines. Engadget quotes Meta's own description of the model as "the most capable US open-weight model to date," a claim the company makes and that the outlet passes along without independent verification.
The two reports frame the launch differently. TechCrunch emphasizes the local-running angle and the Apple partnership. Engadget focuses on the practical limit that the memory requirement places on who can actually use it. Neither report, as summarized here, provides detailed benchmark results, pricing, licensing terms or a release schedule, so those details remain unclear from the available coverage.
The announcement arrives as companies compete to offer capable models that users can download and run themselves, rather than only accessing them through paid online services.
Background
Open-weight models are AI systems whose trained parameters, often called weights, are published so that outside developers, researchers and companies can download them. This is different from a fully open-source release, which typically also includes training code and data. Many open-weight releases carry their own license terms, and those terms can restrict commercial use, redistribution or certain applications, so the label alone does not tell a user what they are allowed to do.
Running a model locally means the computation happens on the user's own hardware rather than on a provider's servers. This can offer more control over data, lower ongoing costs and independence from an internet connection, but it puts the full burden of memory and processing onto the device. Large language models are typically measured by their parameter count, and the memory needed to load them generally grows with that size. Techniques such as quantization, which stores weights at lower numerical precision, can shrink the memory footprint, though often at some cost to output quality. Apple's silicon uses unified memory shared between the processor and graphics components, which makes high-memory configurations a meaningful factor for anyone running large models on a Mac.
Meta has released earlier models under open-weight terms, and its approach has drawn both praise from developers who value access and debate from those concerned about the risks of publicly available powerful models. Apple, for its part, has promoted on-device AI features as part of its product strategy. The intersection of these two trends, a large model aimed at a consumer desktop platform, is the context in which this launch appears.
Why it matters
The most immediate question is accessibility. A model that requires around 1TB of memory is not something most individuals can run on a standard laptop or desktop, even on high-end Macs. That means the practical audience may be limited to research labs, large companies or enthusiasts with exceptional hardware, at least for the full version. Whether smaller or compressed variants are planned is not clear from the coverage available.
The launch also highlights a broader trend in which major technology companies seek to shape the AI ecosystem by releasing capable models that others can build upon. Open-weight releases can accelerate research and give developers more options, while also raising questions about how such models might be used once they are widely available. Each release tends to revive those debates among policymakers, security researchers and the developer community.
For Apple, a partnership that brings a major third-party model onto its hardware could strengthen the case that Macs are serious machines for AI work. For Meta, releasing a capable model openly can build goodwill with developers and expand the reach of its technology, though the company's commercial motives are not spelled out in the reports.
Finally, the memory requirement serves as a reminder that the gap between what AI models demand and what consumer devices provide remains substantial. How quickly that gap closes will shape who gets to run advanced AI locally.
What to watch
Watch for the full license terms, since they will determine how the model can be used, modified and distributed. Look for independent benchmark results that test Meta's claim of being the most capable US open-weight model, as performance comparisons from outside parties will matter more than the company's own description.
Pay attention to whether Meta or Apple releases smaller or quantized versions that reduce the memory requirement to something a typical Mac can handle. Also watch how developers and tool makers respond, and whether the local-running feature becomes a visible part of Apple's software updates.
Viewpoints
Commentary from our columns: three angles on this story.
The Skeptic open questions and weak points
Every launch comes with a headline, and this one has two. The first says Meta's new model runs on a Mac. The second, buried in the details, says you would need roughly 1TB of memory to do it. Those two claims do not sit comfortably together for most readers. A model that only runs on machines almost nobody owns is a demonstration, not a product, and it is fair to ask who the launch is actually for.
The benchmark claim deserves the same scrutiny. Meta calls this the most capable US open-weight model to date, but that is the company describing its own work. Until independent testers publish comparisons under the same conditions, the phrase is a marketing line. Open questions also remain about licensing, since "open" can mean very different things depending on what the terms allow. Readers should treat the excitement as a reason to wait for evidence, not as evidence itself.
The Builder the engineer's view
From an engineering standpoint, the interesting part is the memory story. Unified memory on Apple silicon is a real advantage for large models, because the processor and graphics share one pool. But a terabyte is a number that changes the conversation entirely. It suggests that the model, at least in this release, is being presented at full precision rather than in a compressed form.
The practical next step is the one developers will look for: quantized builds, smaller variants and clear documentation on what quality is lost at each step. Those are the artifacts that turn a headline into something you can actually run. If Apple's toolchain and Meta's weights can be combined into a workflow that handles memory well, that would be a meaningful contribution. Until then, I would treat this as an interesting signal about where the stack is heading, not a tool to install this week.
The Rulebook privacy, safety and regulation
When a powerful model is published with downloadable weights, the rules that govern its use become as important as its speed. Open-weight releases are often described as open, yet their license terms can restrict commercial use, redistribution or certain applications. Readers and developers should read those terms closely before building anything on top of the model.
There is also a question that open releases always raise: once weights are public, they cannot be recalled. That is a reason for careful thought about safeguards, documentation and responsible-use policies, and for regulators to ask what openness means in practice. Users running a model locally may gain privacy, since their data does not have to leave the device, which is a genuine benefit. But that privacy comes with responsibility for how the model is deployed. The coverage does not describe the safeguards or terms in detail, and that gap is where the conversation should focus next.