Meta Muse Spark Open is an 816-billion-parameter model designed to run locally on two powerful Mac Studios. It mainly concerns organizations and demanding users who do not want to send sensitive material to an external cloud service. The release is notable because a model of this scale can now fit on a desk, although it is still a long way from running on an ordinary MacBook.
What is Muse Spark Open?
Muse Spark Open is an open-weight model, meaning its trained weights, the numerical values learned during training, can be downloaded. That does not automatically make the entire system open-source, but it offers more control and transparency than a model available only through someone else’s cloud service. Meta is distributing the weights in BF16 and NVFP4 formats through Hugging Face.
The model has approximately 816 billion parameters in total. It uses a Mixture of Experts (MoE), an architecture in which only selected parts of the model process each input. Eight of its 256 specialized areas are active at a time, bringing the number of parameters used per processed token to about 42 billion. The 816-billion figure therefore describes the model’s total size, not the computing work performed at every step.
Muse Spark Open accepts text, images, audio, and video but produces text as its output. Its context window holds up to one million tokens, the small units into which text is divided. That capacity can accommodate extensive documents or long workflows, although context size by itself says nothing about the reliability of the answers.
Meta calls it the most capable open-weight model from the United States. This is a provider claim and has not been independently verified. The published benchmark results show strong performance in tool use and knowledge tasks, but a more mixed picture in coding tests.
How does such a large model run on Macs?
For local operation, Meta specifies two Mac Studios with M5 Ultra chips and 512GB of unified memory each. Together, that amounts to one terabyte of memory. TechCrunch similarly reports a requirement of at least one terabyte of RAM and identifies the M5 Ultra as the appropriate Mac configuration.
The two computers use MLX, a machine-learning framework for Apple silicon. Reports say Meta and Apple worked together for several months on Mac support. Meta describes Muse Spark Open as the first model in its class that can operate at a desk using two M5 Ultra Mac Studios and the latest version of MLX.
The word “desktop” needs some perspective. Two Mac Studios may be more compact and accessible than a conventional data center, but they remain specialized equipment. The supplied sources provide neither purchase prices nor power-consumption figures, so the total cost cannot be calculated reliably.
The lead report also lists four Blackwell Ultra graphics processors or eight H200 processors as alternatives for running the NVFP4 version. Heavily compressed versions are said to work in setups with about 300GB. The available sources do not independently establish whether those reductions preserve the same quality and speed.
Why does local operation matter?
With a local model, inputs can remain on your own hardware as long as you do not connect additional external services or tools. This can be useful for confidential documents, internal media collections, or material containing personal information. The model can also operate without a continuous connection to the provider once the necessary files and software are in place.
One workplace example comes from Meta. Paired with the Muse Code programming tool, the model worked on a Mastodon client, fixed bugs, added features, and completed a full build. In another example, it operated DaVinci Resolve to review and crop video footage. Both demonstrate long, multistep workflows, but they are provider examples rather than independent field tests.
Hugging Face describes additional input options: the model can examine a screenshot, listen to a voice note, or review a screen recording before generating text. In a support case, for example, it could inspect a recording of a user’s actions; in media production, it could review footage and prepare editing steps. The sources do not establish how reliably these tasks work outside the demonstrated workflows.
Local operation does not solve every privacy problem. A report about the Muse assistant warns that AI agents may receive access to contacts, messages, photographs, passwords, or banking information. Even if processing happens locally, data belonging to friends, customers, or coworkers does not automatically become fair to use. Technical privacy and the consent of affected people are separate issues.
Pros and Cons of Muse Spark Open
Pros:
- Local control – Confidential inputs do not necessarily have to be transferred to a model provider’s cloud service.
- Multiple input types – Text, images, audio, and video can be processed by the same model, even though its output remains text-only.
- Long context – A window of up to one million tokens provides room for extensive documents and prolonged workflows.
- Open weights – Downloadable model weights provide more flexibility than an exclusively closed service.
Cons:
- Extreme hardware requirements – Two M5 Ultra Mac Studios with a combined terabyte of memory are not typical personal equipment.
- Unclear total cost – The sources provide no hardware prices, energy figures, or estimates for ongoing operational work.
- Mixed performance – Meta reports strong tool-use results, while the model trails competitors in some coding comparisons.
- No automatic data safety – Excessive permissions and third-party data remain problematic even when processing happens locally.
The benchmark results illustrate that range. According to Meta, Muse Spark Open scores 89.2 on MCP-Atlas, ahead of Qwen 3.8 and GLM 5.3. On Terminal-Bench 2.1, however, it scores 76.4, behind Kimi K3 at 85.0, GLM 5.3 at 83.9, and Qwen 3.8 at 82.0. THE DECODER also reports that the open model falls well behind Meta’s paid Muse Wonder Full.
The license also comes with conditions. According to THE DECODER, free use requires the label “Built with Muse Spark Open.” Companies with more than $20 million in annual revenue need a separate agreement with Meta if they offer the model as an application programming interface service. An application programming interface (API) makes model functions available to other applications.
What does this mean in practice?
For beginners: Muse Spark Open is not a sensible first step if you simply want to experiment with local Artificial Intelligence (AI). Meta itself positions Muse Glimmer, a smaller dense model from the same family, as the more suitable option for local use. Beginners are better off viewing Muse Spark Open as evidence that large cloud-scale models are gradually moving onto local high-performance systems.
For advanced users: The model becomes relevant when suitable Mac hardware, clearly defined workflows, and operational expertise are already available. It can then process extensive documents, media, and multistep tasks locally. A comparison against existing tools using your own nonsensitive test material will be more informative than relying only on rankings, since most of the published results come from Meta.
The sources provide no Switzerland-specific information about availability or local data-protection requirements. The model description says its training data covers more than 100 languages, but it offers no separate quality results for German, French, Italian, or Romansh. Local processing may be attractive in Swiss workplaces and education, but it does not replace checks of access rights, consent, and internal policies.
Meta and Apple present Muse Spark Open as proof that very large models can operate outside a data center. That is true in a narrow sense: two powerful Mac Studios fit on a desk and provide more local control. The unresolved issues are the real cost, performance beyond provider demonstrations, and the risk that an autonomous system may receive overly broad access to data even while running locally.
Sources
- Meta Releases Muse Spark Open: 816B Open-Weight MoE With 1M Context – MarkTechPost, 2026-10-16
- Meta launches an open-weight AI model that runs on Macs – TechCrunch, 2026-10-15
- Muse Spark Open: Meta meldet sich mit Open-Weight-Modell optimiert für Apples M5 zurück – THE DECODER, 2026-10-15
- Meta and Apple team up to make Muse AI models run better on Macs – The Verge, 2026-10-15
- Welcome Muse Spark Open: open-weight frontier multimodal mixture-of-experts model from Meta – Hugging Face, 2026-10-15
- Friends don’t let friends feed their data to AI agents – The Verge, 2026-10-15
Image: Tranmautritam via Pexels