# Meta's Muse Spark Open Runs Locally on Apple Mac Studios

- URL: https://usa-24.com/meta-muse-spark-open-mac-studio/
- Publisher: USA 24
- Section: Artificial Intelligence (AI)
- Published: 2026-10-19T18:49:23.907Z
- Updated: 2026-10-19T18:49:23.913Z
- Topics: Meta, Open-Weight Models, Apple, Mac Studio, Local AI

> Meta says its 816-billion-parameter Muse Spark Open is the most capable US open-weight model, and it can run on two Apple Mac Studios. The license and memory bill limit who can use it.

## Key takeaways

- Meta released Muse Spark Open, an 816-billion-parameter open-weight model, on Thursday, October 15, 2026.
- Meta's optimized setup uses two M5 Ultra Mac Studios with 512GB of unified memory each, or four 256GB units for 1TB total.
- The Muse Model License allows free use for many developers but limits commercial use, and published summaries of its exact terms conflict.
- Muse Spark Open is derived from Meta's closed Muse Spark, while the full Muse Wonder model stays behind paid Muse Code plans at $50 and $200 a month.
- Meta's claim that Muse Spark Open is the most capable US open-weight model comes from Meta and has not been independently verified.

Meta Muse Spark Open can run on a desk, as long as the desk holds two Apple M5 Ultra Mac Studios. The 816-billion-parameter model became downloadable on Thursday, October 15, and Meta calls it the most capable open-weight model in the United States. The weights are public, but the license terms, the memory bill and the fine print decide who can actually use it.

## What Meta Muse Spark Open released on October 15

Meta posted the weights on [Hugging Face](https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B) and through other distribution partners. The model has 816 billion parameters, with about 42 billion active for each token, and a context window of 1 million tokens. Meta's [developer blog](https://dev.meta.ai/resources/blog/build-with-muse-spark-open) lists support for llama.cpp, SGLang and vLLM, along with cloud and API providers, so teams can self-host or rent capacity.

Meta derived the open model from Muse Spark, its closed flagship. According to [Technology.org](https://www.technology.org/2026/10/16/meta-muse-spark-open-816b-open-weight-mac-studio/), the open version received extra safety post-training so it can resist misuse when it runs without Meta's system-level guardrails. Once the weights leave Meta's servers, no hosting company sits in the middle filtering requests, so that extra training carries more weight than usual.

Meta's own developer post makes the boldest claim: “the most capable US open-weight model running on the desktop.” A Meta AI announcement, as reproduced by the NewsCord aggregator, says the model “brings leading agentic capabilities to your desktop for the first time.” Both are company statements. Meta's benchmarks have not been independently verified, so treat “most capable” as a claim to test, not a settled fact.

Mark Zuckerberg posted on X that the model “runs great on new Mac Studios thanks to work with @johnternus and Apple,” as [Benzinga](https://www.benzinga.com/markets/tech/26/10/62403354/meta-muse-spark-open-weight-apple-mac-studio) reported. That post is the clearest sign of how central Apple's hardware is to the launch.

## Which Mac Studio setups can run Muse Spark Open

The hardware is the real gate. Meta's developer guidance describes two M5 Ultra Mac Studios with 512GB of unified memory each, or four 256GB units for 1TB in total. The setup runs through Apple's MLX framework, which Meta and Apple tuned together.

- Two M5 Ultra Mac Studios with 512GB each is the optimized configuration Meta describes.
- Four 256GB units also add up to 1TB, which is the other configuration Meta lists.
- A compressed 4-bit build still occupies roughly 493GB, per [iClarified](https://www.iclarified.com/102861/meta-releases-muse-spark-open-ai-optimized-for-mac-studio), before the extra memory needed to run it.
- Another report puts GPU memory on servers at about 700GB, a very different deployment from a desk.

[TechCrunch](https://techcrunch.com/2026/10/15/meta-launches-an-open-weight-model-that-runs-on-macs/), citing Engadget, says the model needs at least 1TB of RAM. Both of Meta's configurations total 1TB, so the reports are less far apart than they first look. What the sources leave open is how much memory a long context consumes once the weights are loaded.

Be realistic about who this is for. Two high-memory Mac Studios are an expensive purchase, and most readers will reach Muse Spark Open through hosted providers. The local story still matters as a signal. Apple's M-series chips are being pitched as a way to run frontier-sized models without a GPU cluster, a shift from the typical cloud-first model.

## What the Muse Model License allows

The weights ship under the Muse Model License, not an unrestricted open-source license. iClarified describes free use for many developers with limits on commercial use. CNBC, in its [report on the Apple collaboration](https://www.cnbc.com/2026/10/15/apple-meta-partner-ai.html), calls the terms “limited commercial use.” Another blog describes broad rights to copy, modify, distill and sell, with a $20 million revenue threshold for API sellers. Those accounts do not match.

Read the license text on Hugging Face before you build a product on the model. The gap between free use for many developers and broad commercial rights determines whether a startup can sell a service built on these weights. No news summary, this one included, should settle that question for you.

## How Apple fits into the launch

CNBC reports that Apple and Meta are collaborating to make the model easier to run locally on clustered Mac Studio systems. Representatives for both companies declined to say whether the work is the start of a deeper relationship. Reports point to an Apple event on October 28, but none of the sources confirm that AI will be part of it.

## How Muse Spark Open compares with Muse Wonder

Meta released two models in one week that point in different directions. Muse Wonder launched on October 12 without publicly downloadable weights, and VentureBeat reports that only the weaker of its two models is available through an API. Muse Wonder's API costs $4 per million input tokens, $20 per million output tokens and $0.20 per million cached input tokens. Muse Wonder Full is sold only through paid Muse Code plans at $50 and $200 a month, according to [Technology.org](https://www.technology.org/2026/10/13/meta-muse-wonder-frontier-model-pricing-muse-code/).

That split is the strategic story. VentureBeat notes that Meta's earlier Llama models helped popularize downloadable AI models that organizations could run on their own infrastructure, and that the Muse Spark Open launch marks a significant shift for the company. Meta is keeping its strongest closed model behind paid plans while releasing a large open model to developers. For teams that do not want to send data to a cloud API, the open option is the useful one.

Open weights also sit under policy pressure. China has proposed export curbs on 23 AI technologies, [from models to chips](https://usa-24.com/china-ai-export-controls-draft/), which could affect how labs distribute models abroad. Nvidia's [Nemotron 3.5 Super VL](https://usa-24.com/nvidia-nemotron-3-5-super-vl/), an open 120B model for agents, shows how crowded the open-weight field is becoming. For a look at how a closed rival is structuring access, see [Google's paid Gemini 4 Argon rollout](https://usa-24.com/gemini-4-argon-price-access-2/).

## Practical steps before you download

- Read the Muse Model License on Hugging Face before any commercial use, because published summaries disagree.
- Check the memory numbers against the model card before buying hardware. Compare the 512GB and 256GB configurations with the roughly 493GB 4-bit figure.
- If you use Muse Code, review Contributor mode. [Coursiv reports](https://coursiv.io/blog/muse-wonder) that its default lets Meta train on prompts, code and files in exchange for higher usage limits, which is a separate issue from the open weights. Check your settings before pasting proprietary code or files.
- Run your own tests. Meta's benchmarks are unverified, so evaluate the model on the tasks you actually need done.

## What to watch next

Several open questions should resolve in the coming weeks. Independent evaluators can test the “most capable” claim against other open models, and developers will show how the license works in practice. Apple's reported October 28 event could also reveal whether the Mac Studio collaboration is a one-off or the start of something larger. Hosted pricing for Muse Spark Open is not in the reporting yet, and it will shape whether most readers ever run the model on their own hardware.

## Frequently asked questions

### Can I run Meta Muse Spark Open on a single Mac Studio?

Meta's optimized setup uses two M5 Ultra Mac Studios with 512GB each, or four 256GB units. The sources describe clustered configurations rather than a single machine. A compressed 4-bit build still occupies roughly 493GB before the extra memory needed to run it.

### Can I use Muse Spark Open for commercial products?

The answer depends on which summary you trust. iClarified and CNBC describe limits on commercial use, while another blog describes broader rights with a $20 million revenue threshold for API sellers. Read the Muse Model License on Hugging Face before you ship a product.

### Where can I download Muse Spark Open?

Meta distributes the weights through Hugging Face and other distribution partners. Meta also lists support for llama.cpp, SGLang and vLLM, along with cloud and API providers. Hosted options are available for readers who do not own high-memory hardware.

### Is Muse Wonder open-weight too?

No. Muse Wonder launched on October 12 without publicly downloadable weights. Its API costs $4 per million input tokens and $20 per million output tokens, and the full version is sold only through paid Muse Code plans.

## Sources

1. [Meta Releases Muse Spark Open AI Optimized for Mac Studio — iClarified](https://www.iclarified.com/102861/meta-releases-muse-spark-open-ai-optimized-for-mac-studio)
2. [Apple and Meta team up to make Muse AI model run on Mac computers — CNBC](https://www.cnbc.com/2026/10/15/apple-meta-partner-ai.html)
3. [Meta Releases Muse Spark Open, an 816B-Parameter Open-Weight Model — Technology.org](https://www.technology.org/2026/10/16/meta-muse-spark-open-816b-open-weight-mac-studio/)
4. [Build with Muse Spark Open — Meta AI developer blog](https://dev.meta.ai/resources/blog/build-with-muse-spark-open)
5. [meta-models/Muse-Spark-Open-816B-A42B — Hugging Face](https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B)
6. [Meta launches an open-weight AI model that runs on Macs — TechCrunch](https://techcrunch.com/2026/10/15/meta-launches-an-open-weight-model-that-runs-on-macs/)
7. [Mark Zuckerberg Opens Meta's Massive Muse Spark AI Model — Benzinga](https://www.benzinga.com/markets/tech/26/10/62403354/meta-muse-spark-open-weight-apple-mac-studio)
8. [Meta's new Muse Wonder and Muse Wonder Full models achieve frontier performance, but only the weaker one is available via API — VentureBeat](https://venturebeat.com/technology/metas-new-muse-wonder-and-muse-wonder-full-models-achieve-frontier-performance-but-only-the-weaker-one-is-available-via-api)
9. [Meta Launches Muse Wonder, Its First Frontier AI Model, at $4 per Million Tokens — Technology.org](https://www.technology.org/2026/10/13/meta-muse-wonder-frontier-model-pricing-muse-code/)
10. [Meta Muse Wonder vs Wonder Full: What's New, Who Can Use Each, Pricing, and Benchmarks — Coursiv](https://coursiv.io/blog/muse-wonder)