493 gigabytes of memory, two high-end desktop computers, and a single Thunderbolt cable: that’s what it takes to run one AI model locally today. On October 15, 2026, Meta released the weights for Muse Spark Open, a model that hits exactly these dimensions and that the company calls the most capable US-developed open-weight model released to date. Local operation is only possible thanks to close collaboration with Apple, whose M5 Ultra chips and MLX framework bring the model to consumer hardware, though only when two Mac Studios are paired together.
The release marks a notable moment for the open-source AI scene: a model with 816 billion parameters, until now conceivable only inside data centers, can now run on a desk, albeit with substantial hardware investment. For developers, researchers, and businesses this opens new possibilities, but it also raises questions about practicality, cost, and the credibility of Meta’s own performance claims.
Table of Contents
Image: "Mac Studio (2022) front.jpg" by Yasu, licensed CC BY-SA 3.0, source: Wikimedia Commons
Key Takeaways
- Muse Spark Open has 816 billion total parameters, with a mixture-of-experts architecture keeping only 42 billion active per token.
- The model supports a context window of up to 1 million tokens and natively processes text, images, audio, and video as input.
- Thanks to the Meta-Apple collaboration, the model runs locally on two Mac Studios with M5 Ultra chips, connected via Thunderbolt.
- In 4-bit quantization, Muse Spark Open needs roughly 493 GB of memory — more than a single Mac Studio provides.
- TechCrunch pointed out that the “most capable open-weight model” claim rests entirely on benchmark data Meta published itself.
What Is Muse Spark Open?
Muse Spark Open is Meta’s newest open-weight language model, with weights made publicly available since October 15, 2026. According to Meta, it is the most capable open-weight model developed in the US to date. That assessment comes directly from the company itself and, as TechCrunch reported, has not yet been independently verified by third parties.
The model is specifically designed for demanding technical use cases. According to Meta, these primarily include:
- Coding tasks, including complex software engineering
- Agentic workflows, where the model independently plans and executes multiple steps
- Long-horizon tasks that span long time periods or many intermediate steps
- Multi-step reasoning across multiple levels of logical inference
This focus sets Muse Spark Open apart from pure chat models and positions it more firmly in the professional developer-tooling space.
The Technical Architecture in Detail
At its core, Muse Spark Open uses a mixture-of-experts (MoE) architecture. This design principle allows building a very large model without requiring the entire network to be active for every single request.
Concretely, the structure looks like this:
| Spec | Value |
|---|---|
| Total parameters | 816 billion |
| Active parameters per token | 42 billion |
| Experts per layer | 256 |
| Active experts per token | 8 |
| Maximum context window | 1 million tokens |
| Input modalities | Text, images, audio, video |
| Output modality | Text |
Because of the mixture-of-experts structure, only eight of 256 available experts per layer are activated per token. That significantly reduces the actual compute cost compared to a densely connected model of the same total size. Memory requirements still remain high, though, since all 816 billion parameters must be held in memory even when only a fraction of them actually compute per request.
The context window of up to one million tokens is among the largest currently available for open-weight models. In practice, this means Muse Spark Open can process extremely long documents, large codebases, or hours-long conversation histories at once without losing context.
Native multimodality is particularly worth highlighting. Unlike models retrofitted to handle images or audio, Muse Spark Open was built from the ground up to process text, images, audio, and video as input. Output remains text-only.
The Apple Partnership: Muse Spark Open on Mac Studio
Perhaps the most remarkable aspect of this release isn’t the model itself, but how it can be run locally. Meta optimized Muse Spark Open together with Apple for Apple’s own MLX framework. Meta CEO Mark Zuckerberg publicly thanked Apple’s hardware chief John Ternus for the collaboration.
“The most capable US-developed open-weight model now runs locally on Apple consumer hardware” sums up the core message of this partnership.
This partnership matters because it connects two worlds that had previously operated separately: Meta’s large-language-model research and Apple’s hardware ecosystem built around its M-series chips.
Why Two Mac Studios Are Needed
Apple’s technical blog post, published via Apple Machine Learning Research, provides the key memory details. In 4-bit quantization, a common method for compressing model weights and saving storage, Muse Spark Open requires roughly 493 GB of memory.
That’s the central problem: a single Mac Studio with an M5 Ultra chip offers, according to Apple, up to 512 GB of unified memory. In theory, the model would just barely fit. In practice, though, the operating system, inference software, and other processes consume additional memory, so a single Mac Studio isn’t enough.
The solution: two Mac Studios are connected via a Thunderbolt cable. Using tensor parallelization, a technique that splits the model’s computations across multiple machines, the two computers share the load. Each machine handles part of the tensor operations while results are synchronized over the Thunderbolt connection.
M5 Ultra Performance Numbers
The technical specs behind this setup are considerable:
- Up to 512 GB of unified memory per Mac Studio
- 1.2 TB/s of memory bandwidth
- With four paired Mac Studios: over 36 tokens per second
- With two Mac Studios using tensor parallelization: roughly 25 tokens per second
These numbers show a clear tradeoff: more paired machines mean higher speed, but also significantly higher hardware costs. The two-Mac-Studio configuration represents a practical minimum to get Muse Spark Open running locally at all, while the four-machine setup is likely intended for more demanding, latency-sensitive use cases.
For developers, this means: anyone seriously planning to run Muse Spark Open locally needs to invest in at least two high-end Mac Studios — a substantial financial hurdle compared to simply downloading the model weights.
Licensing and Availability of Muse Spark Open
Muse Spark Open is distributed under the so-called “Muse Model License”. This is not a fully free open-source license, but one with limited commercial usage terms. Companies and developers planning commercial use of the model should carefully review the exact license terms before building products on top of it.
Despite the licensing restrictions, availability is broad. The model weights are available for download or use through the following platforms:
- Hugging Face
- Amazon SageMaker AI
- Google Cloud Model Garden
- Microsoft Foundry Models
Several specialized infrastructure providers also offer access to Muse Spark Open, including:
- Together AI
- Fireworks
- Baseten
- Modal
- Groq
- Nebius
- OpenRouter
This broad distribution across established cloud and inference platforms makes it easy for developers and companies to try the model without immediately investing in their own hardware. Anyone who values fully local, privacy-friendly inference, though, currently has little choice but the Mac Studio route.
Critical Assessment: How Capable Is Muse Spark Open Really?
Meta’s claim that Muse Spark Open is the most capable US-developed open-weight model deserves closer scrutiny. TechCrunch’s coverage noted critically that this classification rests on benchmark results Meta published itself. Independent, third-party testing is not yet available.
This limitation isn’t unusual in the AI industry — many vendors initially present their own models alongside self-collected comparison figures. Still, it’s important for technically minded readers to understand the distinction: a benchmark result coming from the manufacturer itself can be methodologically designed to cast that company’s own model in an especially favorable light. Only independent retesting by research institutions, academic groups, or competing companies can show whether the performance claims hold up in practice.
For developers, this means: before putting Muse Spark Open into production, it’s worth running your own application-specific tests. The claimed strengths in coding, agentic tasks, and long-context handling should be validated against real project requirements rather than relying solely on the official figures.
Where This Fits in the Open-Weight Landscape
Muse Spark Open joins a dense string of AI model releases in October 2026, including Google’s Gemini 4 Argon, which underwent a hands-on review around the same time.
With Muse Spark Open, Meta continues its strategy of making powerful models openly accessible rather than offering them exclusively through closed APIs. The close technical cooperation with Apple adds a new angle: instead of relying solely on cloud infrastructure, Meta is specifically investing in optimization for consumer and prosumer hardware.
This development could prove influential. If other vendors pursue similar partnerships between model development and hardware optimization, access to extremely large models could gradually shift from pure data centers toward high-performance workstations, even if, for now, only a small group of technically and financially well-equipped users can take advantage of it.
Conclusion and Next Steps
Muse Spark Open marks a significant, if nuanced, milestone. The technical specs — 816 billion parameters, a mixture-of-experts architecture, a one-million-token context window, and native multimodality — describe an ambitious model purpose-built for demanding coding and reasoning tasks. The Apple collaboration also makes it possible, for the first time, to run a model like this outside classic data-center infrastructure, albeit with substantial hardware investment in the form of two paired Mac Studios.
At the same time, the question of how it actually performs against competing models remains open as long as independent benchmark results are missing. Interested developers and businesses should consider the following steps:
- Review the license terms: Read the “Muse Model License” carefully before any commercial deployment.
- Run your own tests: Use Meta’s benchmark results as a starting point, but verify them with your own project-specific tests.
- Calculate hardware costs: Weigh investing in two Mac Studios with M5 Ultra chips against cloud-based alternatives like Together AI or Groq.
- Compare available platforms: Depending on existing infrastructure, access via Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, or Microsoft Foundry Models may make more sense than running it locally.
Muse Spark Open points toward where open-weight AI development could be heading: ever larger, more capable models that, through close hardware partnerships, increasingly become usable outside the cloud as well. Anyone active in this space should keep a close eye on further developments, especially independent benchmark results.
