Meta released Muse Spark Open on October 15, and you can download the weights. Meta's demo ran it on two Apple M5 Ultra Mac Studios, which, in the company's words, "behave like one supercomputer with a terabyte of unified memory." So "AI on your desktop" is doing quite a lot of work here.
Still, a company can now run the model on hardware it owns. If your team handles sensitive work it wants to keep in-house, that's worth a look. If you're a freelancer who needs occasional help with a document, setting this up may be more work than the problem it solves.
What Meta released
Meta describes Spark Open as a multimodal model built for coding and agent tasks. It has 816 billion total parameters, 42 billion active parameters and a one-million-token context window. Hosted providers offer it too, so you don't have to download it to try it.
Those two parameter counts explain the hardware bill. Spark Open is a mixture-of-experts model: for each token it uses only selected parts of its network, which cuts the computation needed at that moment. The rest of the weights still have to be stored and ready. When you plan memory, budget for the whole model, not the 42 billion.
Meta's demos include changing an iOS app and editing footage in DaVinci Resolve. They show the kind of work Meta is aiming at. They don't tell you how fast your own jobs will finish or how much fixing they'll need.
Don't confuse it with the Muse Wonder and Wonder Full models we covered earlier this week. Those are closed models, available through Meta's own products and services. Spark Open is a separate, downloadable release, so this isn't Meta publishing Wonder Full's weights, and sharing the Muse name doesn't mean the models can do the same things. If you compare results, write down exactly which model you used and what it ran on.
The license has a $20 million line
The model license lets you modify and fine-tune Spark Open, subject to its conditions and usage policy, and products must show the required attribution. The line is for companies selling the model as a service. If your combined revenue with affiliates tops $20 million over any twelve consecutive months, you need a separate agreement with Meta.
The license expressly leaves internal use, and ordinary product features that aren't a model-as-a-service business, outside that threshold. If you're thinking about an internal document assistant, that matters. It's also why "open weights" is the more accurate label: not every kind of commercial use gets the same permission.
Owning the hardware changes the bill
Running it locally can make sense if your workload is steady or you need tighter control over sensitive inputs. It also means a big upfront purchase and an ongoing job. Someone has to install updates, manage access and keep the thing running when colleagues depend on it.
To compare, take a week of typical work. Count the results that were actually acceptable and the time people spent fixing them. Add setup and maintenance to the local option and usage charges to the hosted one. A smaller monthly invoice isn't much of a saving if a colleague spends their Fridays repairing the server.
And check the tools, not just the model. A locally running agent can still send a document to an outside search or storage service if you connect one. Look at those connections before you call the whole setup private.
For most readers, the sensible first step is to try Spark Open through a hosted provider on a few tasks you know well. The supercomputer can wait until the results give you a reason to buy one.
Cover: AI-generated editorial illustration by TMRW using the Meta logo. The hardware scene is illustrative.
