What happened
On October 15, 2026, Meta released the weights of Muse Spark Open. Meta describes it as the most capable US open-weight model to date. It is a multimodal mixture-of-experts model with 816 billion total parameters, 42 billion of them active for each token, and a 1 million token context window. Meta built it for coding and agentic work. Meta's post says it runs on two Apple M5 Ultra Mac Studios using Apple's MLX framework, because each machine can use up to 512 GB of unified memory. CNBC reports that Apple and Meta worked together on this, that customers can pair two to four Mac Studios, and that a Mac Studio with the required M5 Ultra chip and 256 GB of memory costs at least $10,799. A 512 GB version is due later this month. CNBC also notes that users would need to buy more than one machine. The weights are available through Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry. Seven hosted providers also offer it through an API: Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI. The model card says that running the weights on your own system takes a GPU setup with at least 700 GB of video memory for the 4-bit version, or 2 TB for the full-precision version. The license is the Muse Model License, which Meta calls "limited commercial terms." It lets many users run the model for free, but it is not a standard open-source license.
Why it matters for business owners
Until now, the largest and most capable models mostly lived behind someone else's API. If you wanted a model of this size, your prompts and files went to a vendor. Meta's release means a model in that class can, in principle, be run on hardware you own, in an office or a rack, with nothing leaving the building. That speaks directly to the three things owners ask about private AI: where the data goes, what the ongoing bill looks like, and who can change the terms on you. Running a model yourself changes the answer to all three. But the announcement is pitched at developers. Meta's AI chief Alexandr Wang told CNBC he expects companies and developers to "tinker with it at small scale" and then scale up. That is a fair description of where most businesses should start too.
What owners should not misunderstand
Do not read "runs on a desktop" as "runs on a laptop." This is two or more high-end workstations with a combined memory of around a terabyte. The starting price CNBC quotes is for one machine with less memory than the setup Meta describes, and the larger version had no published price in the reporting. Do not assume local means free. Hardware, power, setup, updates, and someone who can keep it running are real costs. Apple has pitched the idea of no cost per token. That is true of the metering and false of the total bill. Our earlier piece on Apple's no-cost-per-token pitch covers how to check that math. Do not trust the launch benchmarks as a buying signal. The performance figures are Meta's own, and the full evaluation report was described as coming later. A model that scores well on a coding test may not be the right model for your quoting, support, or document work. Do not assume open weights means no strings. The license asks products built on the model to display "Built with Muse Spark Open." Companies with more than $20 million in gross revenue that sell access to the model as a service need a separate written agreement with Meta. Meta says the clause does not apply to internal use or to features in your own product. Reading the license, and asking a lawyer where your use lands, is part of the decision. And do not skip the safety side. Meta says it added safety training for the open release. Once weights are on your hardware, access control, logging, and what the model is allowed to touch are your job, not Meta's.
The operational lesson
A model this size creates three separate choices, and owners often blur them. First, use the model through a hosted provider. You pay per use, nothing needs buying, and your data goes to that provider under its terms. This is the cheapest way to find out whether the model is any good for you. Second, rent capacity in your own cloud account. The weights run in an environment you control, such as a private endpoint in a cloud you already use. You keep tighter control than with a shared API, without buying hardware. Third, run it on your own machines. This gives the strongest data boundary and the most control, and it carries the highest fixed cost and the most work. The model is the same in all three. The data exposure, the cost shape, and the staffing needed are not. The right order for almost everyone is to test first, rent second if the test holds up and the data needs it, and buy last. Hardware should follow a workload that is already proven, not a headline.
What a serious business should do next
Write down the one or two tasks you would want a model like this to do. Be specific: for example, drafting replies to a certain type of customer email, or checking contracts against a checklist. If you cannot name the task, there is nothing to test yet. Run those tasks through a hosted version using non-sensitive or sample material, next to the tool you use today. Compare the results on your own work, not on a chart. Note the cost per finished task, not the price per million tokens. If the model holds up, ask where your data may go. For sensitive work, price out a private endpoint in a cloud you already use before pricing hardware. Then compare it against the hosted bill at your real volume. Read the Muse Model License before building anything customer-facing on it, and have a qualified professional confirm how it applies to you. Nothing here is legal advice. If you do decide to buy machines, plan the people side first: who sets it up, who patches it, who gets access, and what happens if one box fails. And if your current setup already does the job, leave it alone. A bigger model is not a reason to change a workflow that works.
The Atlacis view
Atlacis helps owners slow down before an AI decision, map the workflow, and match the tool to the job, including whether private AI is worth it at all. A very large open model is interesting, but the useful question is rarely how big it is. It is whether your data, volume, and risk justify running anything yourself. If you are weighing hosted, rented, or owned AI and want a second view before you spend on hardware, that is a good conversation to have first.
The short version
- On October 15, 2026, Meta released the weights of Muse Spark Open (816B parameters, 42B active) and said it runs on two Apple M5 Ultra Mac Studios.
- CNBC reports a Mac Studio with the required chip and 256 GB of memory costs at least $10,799, and you need more than one machine. Self-hosting on GPUs takes at least 700 GB of video memory for the 4-bit version, per the model card.
- The license asks for a "Built with Muse Spark Open" notice and a separate Meta agreement for large companies that sell the model as a service. It is not a standard open-source license.
- Benchmarks are Meta's own. Test the model on your tasks before it changes a decision.
- Test hosted first, rent private capacity second if the data needs it, and buy hardware last.
Where ATLACIS can help
Sources
- Meta AI Research: Introducing Muse Spark Open (October 15, 2026)
- Hugging Face: Muse Spark Open 816B-A42B model card and Muse Model License
- CNBC: Apple and Meta team up to make Muse AI model run on Mac computers (October 15, 2026)
- TechCrunch: Meta launches an open-weight AI model that runs on Macs (October 15, 2026)