Meta's new open-weight model runs on Macs, if you have 1TB of RAM to spare

Meta just dropped what it's calling the most capable open-weight model available in the US, and it comes with an unusual design target: your Mac. The model, Muse Spark Open, is the latest entry in Meta's Muse line and it's explicitly built to run locally on Apple silicon, a pairing that hasn't really been a priority for frontier-scale open-weight releases until now.

The specs are big. Muse Spark Open runs on 816 billion parameters with a 1 million token context window, which puts it in rarefied territory for openly available weights. Meta says its internal evals show the model outperforming other popular open-weight competitors, though as always with vendor-published benchmarks, independent verification will matter once third parties get their hands on it.

Here's the catch, and it's a big one: running this thing locally requires at least 1 terabyte of RAM, according to Engadget. That's not a typo. We're talking about hardware that basically only exists in Apple's highest-end Mac Studio configuration, the one running the M5 Ultra chip. For context, even heavily loaded consumer workstations rarely cross 256GB of unified memory. A full terabyte puts this squarely in "specialized workstation" territory rather than "something a developer runs on their laptop."

That's an interesting tension for an open-weight release. Part of the appeal of open weights is usually accessibility, the ability for a solo developer or small team to download something and run it on hardware they already own. Muse Spark Open technically clears the "open" bar since the weights are available, but the practical bar for actually running it locally is set extremely high. It reads less like a model for hobbyists and more like a flagship showcase for what Apple silicon can do when you throw enough memory at it.

Built for coding and agentic workflows

Muse Spark Open is natively multimodal, meaning it handles multiple input types (text, presumably images and other modalities) without bolted-on adapters. Meta is positioning it specifically for complex coding tasks and agentic workflows, the kind of multi-step, tool-using tasks where a model needs to plan, execute, and self-correct rather than just answer a single prompt.

Meta paired the model with Muse Code, its AI coding terminal, to demonstrate what it can do in practice:

• It handled a full development cycle on a Mastodon client, including bug fixes, new feature implementation, and a complete build
• It was deployed inside DaVinci Resolve, the professional video editing suite, where it reviewed and cropped footage autonomously

That second example is worth pausing on. Getting a language model to meaningfully operate inside a video editing tool, not just generate code that touches the software, but actually review and crop footage, suggests Meta is testing these models as agents that interact with existing professional software rather than standalone chat interfaces. That's a meaningfully different use case than typical coding benchmarks, and it points toward where Meta wants agentic AI to go: embedded inside the tools professionals already use, not replacing them with a new interface.

The timing is also notable. This launch comes just days after Meta released Muse Wonder, described as the company's first frontier model. Two Muse releases in a single week signals Meta is pushing hard on multiple fronts at once: a frontier closed-ish flagship on one side, and a specialized, locally-runnable open-weight model on the other. It's a split strategy that lets Meta compete for both the "best model" headlines and the "most deployable" use case simultaneously.

The open-weight angle matters beyond just licensing terms. Open-weight models typically offer real cost advantages over closed, API-gated competitors since you're not paying per-token inference fees to a third party once you've got the hardware. They also offer transparency, since researchers and developers can inspect the weights directly, and flexibility, since you can fine-tune or modify the model for specific use cases without waiting on a vendor's roadmap. Those are genuine structural advantages over proprietary models from OpenAI, Anthropic, or Google, even if Muse Spark Open's hardware requirements blunt the "run it on your own machine easily" part of that pitch for most people.

Meta says the model is available now through multiple API platforms, which is probably how most developers will actually access it in practice. Unless you're sitting on a maxed-out Mac Studio, hitting an API endpoint is going to be the realistic path for the vast majority of teams who want to experiment with Muse Spark Open without buying new hardware.

What to Watch

The real test here isn't the benchmark numbers Meta published Thursday, it's what independent developers find once they actually load this thing up. Watch for three things: whether third-party evals confirm Meta's claims about beating other open-weight models, whether the 1TB RAM requirement actually gets met in practice by enough Mac Studio owners to generate meaningful usage data, and whether Apple leans further into this kind of partnership. If Apple silicon becomes a genuine target platform for frontier-scale open-weight models rather than an afterthought, that's a bigger story than any single model release, and it's one worth tracking as more labs decide where their weights are optimized to run.