Running a frontier-class AI model on your own desk used to sound like a joke. On Thursday, Meta made it sound almost reasonable. The company launched Muse Spark Open, an open-weight model it calls the most capable of its kind in the United States, and it was designed from the start to run locally on Apple’s computers. The catch is a hardware shopping list that most people will read twice.
The short version
- 816 billion parameters and a 1 million token context window
- Natively multimodal, built for coding and agentic work
- Designed to run locally on Apple silicon, best on a Mac Studio with the M5 Ultra chip
- Needs at least 1 TB of RAM to run locally, according to Engadget
- Available now through several API platforms for everyone else
What Meta actually released
Muse Spark Open is a large, natively multimodal model, which means it handles text and images as part of the same system rather than through bolted-on add-ons. Meta says evaluation results it published show the model beating other popular open-weight options. As always with launch-day benchmarks, those numbers come from the company itself, so treat them as a claim until independent testers repeat them.
What stands out is the target. This is not pitched as a chatbot. Meta describes it as a workhorse for complex coding tasks and agentic labor, meaning jobs where the model plans, uses tools and finishes a multi-step task on its own. To show that, Meta paired it with Muse Code, its AI coding terminal, and ran it through a few real workflows.
Two demos worth noting
- A Mastodon client. The model handled coding fixes for a Mastodon client, including bug fixes, new features and a full build.
- DaVinci Resolve. Meta says the model drove the high-end video editing software, reviewing and cropping footage.
The second demo is the more interesting one. Driving a desktop app is a different skill from writing code, and doing it on a local machine means the footage never has to leave your computer.
The hardware reality check
Here is where the excitement meets a spec sheet. Running the model locally requires at least a terabyte of RAM, per Engadget, and Meta points to the Mac Studio with the M5 Ultra as the best way to do it. That is a very expensive desktop by any standard.
The memory requirement makes sense once you do some rough math. A model’s weights have to sit in memory, and the footprint depends on how tightly they are compressed.
Even the smallest estimate lands in the hundreds of gigabytes before you add the memory a 1 million token context needs while it works. That is why Meta’s 1 TB figure is a floor, not a comfortable target.
| Detail | Muse Spark Open |
|---|---|
| Parameters | 816 billion |
| Context window | 1 million tokens |
| Input types | Natively multimodal |
| Best local hardware | Mac Studio with M5 Ultra |
| Minimum RAM to run locally | 1 TB (per Engadget) |
| Paired tool | Muse Code, Meta’s AI coding terminal |
| Access without the hardware | Several API platforms, available now |
Why open weights still matter
Open-weight models give you something closed models never will: the weights themselves. That brings real advantages, according to TechCrunch’s reporting. They can be significantly more cost-efficient, and they offer a level of transparency and flexibility that proprietary systems do not. You can host them where you like, inspect how they behave and build on top of them without asking permission each time.
For a company that handles sensitive code, medical records or unreleased footage, running the model in-house is the difference between a usable tool and a compliance headache. Even if you never buy a terabyte of RAM, a model like this can be hosted by a cloud provider or a service you trust, which is why the launch through multiple API platforms matters just as much as the Mac angle.
How this fits Meta’s strange year
Meta’s model strategy has zigzagged. Earlier this year it surprised people when it quietly put a 30B model on Hugging Face after saying open source had been a mistake. Just days earlier it launched its first frontier model, which we covered when Meta finally shipped a frontier model. Muse Spark Open is the open counterpart, and at 816 billion parameters it is roughly 27 times larger than that 30B release.
Pairing the launch with Apple is the other notable move. Apple has spent years selling the idea that serious AI should run on your own device, and a model tuned for its chips gives that pitch a heavyweight example. It also lines up with the broader shift toward software that writes software, a trend visible when Google Play started letting people run private, vibe-coded apps on their own phones.
Who should care
- Developers and studios with big budgets. A Mac Studio loaded with memory is pricey, but it is far cheaper than a rack of datacenter GPUs and keeps everything on site
- Teams that cannot send data to the cloud. Local inference keeps code, footage and documents inside your walls
- Everyone else. Use the API route and judge the model on results, not on the hardware story
What to watch next
- Independent benchmarks. Meta’s own comparison is a starting point, not a verdict
- Real memory tricks. Community compression work could shrink that 1 TB floor, as it has for earlier open models
- Licensing details. What you are allowed to build and sell with the weights will decide how useful the “open” label really is
The takeaway is simple. Muse Spark Open proves that a frontier-scale open model can live on a single desktop, but only for people willing to spend serious money on memory. For the rest of us, the real news is that a very large open-weight model is now just an API call away.
Sources and further reading
About this article: GeekBlog covers U.S. technology news, AI, phones, smartwatches and gaming. Every story is written and checked under our Editorial Policy. Spotted a mistake or have a story tip? Contact our editors.
