Is Muse Spark Open on Mac worth it for business admin?

What did Meta actually release with Muse Spark Open?

On October 15, 2026, Meta released Muse Spark Open, an open-weight AI model it calls the most capable US open-weight model to date. Meta also worked with Apple so the model can run on a desktop setup of two M5 Ultra Mac Studios using Apple's MLX framework. That's the headline. The fine print is where small business owners should slow down. Open-weight means you can download the model's trained weights and run them on your own hardware instead of renting access through someone else's chat app. It doesn't mean the model is small or simple, and it doesn't work out of the box. It's more like being handed a world-class engine with no car built around it.

Muse Spark Open is a mixture-of-experts model with 816 billion parameters in total. Only about 42 billion of them are active for any given piece of text. Its context window is one million tokens. That's enough to hold a thick bundle of contracts or a year of meeting notes in a single prompt. It accepts text, images, video and audio, and it answers in text. Meta is pitching it at coding and agentic work, meaning tasks where a model plans several steps and calls tools along the way. The weights come in two official versions: full BF16 precision and a compressed 4-bit NVFP4 edition. Both downloads are gated, so you'll hand over your legal name, date of birth and organization before you get the files.

The model comes under Meta's own Muse Model License, not a standard open-source license like Apache 2.0 or MIT. You can use, copy, fine-tune and build on it for free, but some commercial uses are limited. If you plan to build it into a product, read the terms before you ship anything. Don't mix it up with the consumer assistant, either. The Muse chat app is a separate product that now has its own iPad and Mac apps, and it's still limited by region. Meta also made Spark Open available through the big cloud platforms and several hosted providers. That matters: if you only want to use the model, you don't need a Mac at all.

What does it take to run Muse Spark Open on a Mac?

Running Muse Spark Open on a Mac takes two Mac Studios with the M5 Ultra chip and 512GB of unified memory each. The two machines are linked over Thunderbolt 5 and run the latest version of Apple's MLX framework. Meta also describes a second option with four Mac Studios at 256GB each. A single machine won't do. Even the compressed 4-bit version takes up roughly 493GB before you count the extra memory the model needs while it's working, so it won't fit on one 512GB Mac. The newest MLX release splits the model across the connected Macs and uses the M5's built-in neural accelerators. That's what makes it possible to run the model on desktop hardware at all.

The speed is decent but not instant. In Apple's own tests of the 4-bit model on two Macs, it generated roughly 16 tokens per second with one way of splitting the work and about 25 tokens per second with another. A token is roughly three-quarters of an English word, so replies stream in at about reading pace. That's fine for drafting an email or summarizing a contract. It's less fine if you hoped to get through hundreds of documents over lunch. Those numbers also come from a tuned setup with a fixed, moderate prompt length. Real work with long documents will behave differently, so treat them as a best case, not a promise.

Then there's the cost no benchmark measures: your time. You have to request access to the weights, download hundreds of gigabytes, install MLX and its tools, get the two machines talking to each other and keep everything updated as new releases come out. Most of that happens in a terminal. If you've never typed a command into one, this isn't a weekend project. It's a new hobby. The 512GB M5 Ultra configuration is also the top of Apple's Mac Studio range, and you need two of them. We won't quote a price because it varies by country and configuration. Check Apple's store before you get too attached to the idea.

What do you actually get once Muse Spark Open is running?

Once Muse Spark Open is running on your Macs, you have a very capable raw model that answers prompts, and that's all you have. Nothing fires a reminder at 10 a.m. on the 15th. There's no to-do list, no memory of what you said last week and no way to reach it from WhatsApp. On its own, a model only speaks when spoken to. It can't tap you on the shoulder because it has no clock, no scheduler and no connection to your phone. Every part that turns a model into an assistant has to be built around it. When the model lives on your own hardware, you're the one who builds it.

Take a simple request: 'Remind me to send the VAT report every 15th at 10.' With a ready-made assistant, that's one message. With a self-hosted model, it's a project. You'd need an app or page to type into, a database to store the reminder and a scheduler that wakes up on the 15th. You'd also need a way to send a notification to your phone, plus something that restarts everything when the Macs reboot after an update. Then you'd want it to remember that your accountant is called Dana, read the PDF invoices you forward and check your Google Calendar. Each of those is another piece of software to connect, test and maintain, and each one can break in its own way.

Our honest view: for admin, the model is the least important part. Plenty of models today can write a polite follow-up or find the deadline in a contract. What turns a clever demo into a useful assistant is the boring stuff. The reminder actually goes off, the follow-up actually gets sent and the note turns up again three months later. Muse Spark Open is an impressive release and a big deal for developers who build those boring parts. The same trade-off comes up with other agents that need a lot of setup. If you just want the VAT report sent on time, buying two flagship Macs to host a model is like buying a printing press because you need business cards.

When does local AI make sense for a small business?

Local AI makes sense for a small business when three things are true. Your data can't leave your premises for legal or contractual reasons, someone technical works in-house, and you'll use the model heavily enough to justify the hardware. A law firm with strict client-confidentiality rules might tick all three. So might a clinic handling sensitive records or a software studio building its own AI features. For them, two Mac Studios humming in a back office can be a sensible investment. They pay no per-request fees, no data goes to a third party, and they decide which model version runs and when it changes. That's a real advantage, and we don't want to wave it away.

Most small businesses don't tick those boxes, though. A solo bookkeeper, a plumbing company with three vans, a five-person design studio or a team lead juggling standups usually needs reliable admin, not a private research lab. Their problem isn't a weak AI. They forget to chase the supplier, miss the rent reminder in the group chat or lose track of who promised what. A local model does nothing about any of that until someone builds the missing parts. And if the person who set it up leaves, or just gets busy, the system slowly falls apart. Hardware you own is also hardware you have to fix, usually on the day you're busiest.

Many people overlook a middle path. If the model itself is what draws you to Muse Spark Open, you can use it through a hosted provider without buying any hardware. You can also drop it the moment something better comes along, which at today's pace could be a few months away. If privacy is what draws you, ask a sharper question: what exactly are you protecting, and from whom? A reminder about a dentist appointment or a note that the courier comes on Thursdays rarely needs a private server. Sensitive contracts might. Sort your admin by how sensitive it is, and you'll often find only a small part needs special handling. The rest just needs to get done.

How does a WhatsApp assistant compare for everyday admin?

For everyday admin, an assistant you text in WhatsApp beats a local model on setup, upfront cost and the things that matter day to day. Reminders fire, follow-ups happen and documents get read. With Agent Sonic, there's nothing to install and no hardware to buy. You message it like any other contact. Send 'Remind me to send the VAT report every 15th at 10' and it sets a recurring reminder, and the reminder arrives right in your WhatsApp chat, the one you already check fifty times a day. Send 'Note that Dana's invoice terms are 30 days' and Sonic remembers it. Weeks later, ask 'What were Dana's payment terms?' and you get the answer. No terminal, no scheduler, no second Mac.

Sonic handles the rest of the admin pile too. Forward a PDF contract and ask, 'What are the deadlines, and is anything risky in here?' Send a spreadsheet and ask for the totals by month. Add Sonic to your team's WhatsApp group and it can post a catch-up recap, run a poll, keep a shared list, split expenses or send automatic reminders for standups and timesheets. It can add a meeting to your Google Calendar or log a row in a Google Sheet. On the Pro plan, a Beta feature messages people for you, for example 'Ask the plumber if Thursday still works and let me know.' Sonic shows you the exact message first and sends it from its own number as your assistant. Then it reports the answer back to you.

To be fair about the limits: Sonic isn't a model you own, and it won't run offline on a Mac in your office. If your data has to stay on hardware you control, a local setup is the right choice, and you should budget for the engineering around it. The messaging feature is deliberately limited, too. It's one task at a time, up to three people a day, one-to-one chats only, and because it's in Beta it may be paused from time to time. Curious how other assistants compare? Here's which AI agents you can text on WhatsApp. If you'd rather have the admin handled than manage an AI project, you can get Sonic's number and start today. The plans and pricing are on the website.

Frequently asked questions