Apple and Meta have spent years competing over mobile platforms, privacy, and the future of computing. Their latest move is unexpected: the two companies have worked together to make Meta's enormous new open-weight AI model run locally on Apple's high-end desktop hardware.
The model is called Muse Spark Open. Released on October 15, 2026, it has 816 billion total parameters and is designed for complex coding, multimodal understanding, and AI agents that carry out multistep tasks.
There is one important catch: running it locally takes much more than an ordinary Mac. Apple's demonstration uses multiple high-memory Mac Studio systems connected together, turning this announcement into a story about memory architecture, AI economics, and the growing market for open-weight models.
Conceptual illustration of local AI inference across multiple Mac Studio computers; not an official product photograph.
What Is Muse Spark Open?
Muse Spark Open is Meta's open-weight model for coding, reasoning, and agentic tasks. Meta says it can handle longer workflows, complex software engineering, and tasks that require planning and tool use rather than simply answering a single question.
The model is natively multimodal: it can accept text, images, audio, and video, while generating text as output. Its one-million-token context window is substantial, although it does not guarantee that every detail in a long prompt will be understood perfectly or processed at the same speed.
Meta describes the release as its most capable US-developed open-weight model to date. That is the company's characterization, not an independently established ranking across every AI benchmark.
Why Apple and Meta Are Working Together
The surprising part is not just the model's size. Apple and Meta have historically been rivals, particularly around mobile platforms, app distribution, and privacy. Their current collaboration shows how practical engineering needs can temporarily align companies that still compete in other areas.
Apple contributes MLX, its open-source machine-learning framework designed for Apple silicon's unified memory architecture. Meta contributes the model weights and model-specific optimization work, allowing developers to run the model locally rather than relying exclusively on a remote inference API.
“We're excited for mommy and daddy to be playing nice again.”
Alexandr Wang, Meta's AI chief, commenting on the unlikely partnership. Source: Axios.
The collaboration is specific. Neither company's announcement establishes that Muse Spark Open will become the default assistant on iPhone, replace Siri, or be integrated into every Mac. Those would be separate product decisions.
“Congrats to the team! Amazing to run a model of Muse Spark Open's caliber and scale this efficiently, right from your desk. Mac Studio keeps pushing the boundaries of on-device AI!”
John Ternus, Apple CEO · October 15, 2026
Open the original post to view its live engagement and any attached media. This lightweight link avoids loading a third-party social widget on every page view.How Can an 816-Billion-Parameter Model Run on a Mac?
The key is a technique called mixture of experts, or MoE. Muse Spark Open has 816 billion parameters in total, but Meta lists 42 billion as active per token, meaning the model activates only a subset of its experts for each generation step.
That reduces the computation required for each token, but it does not eliminate the need to store the model's weights. Memory capacity remains the crucial constraint.
Apple's technical report explains that the model would require more than 1.6 terabytes of memory for its parameters alone in its native BF16 precision. With four-bit NVFP4 quantization, the reported model footprint falls to about 493GB, but intermediate inference computations still prevent it from fitting on a single 512GB Mac Studio.
What makes the Mac Studio cluster work?
- Unified memory: The CPU and integrated GPU share a large memory pool, rather than relying solely on a separate graphics card's VRAM.
- Quantization: The four-bit NVFP4 format reduces the memory needed to hold the model's weights.
- MLX distributed inference: The model is split across machines instead of being loaded entirely into one computer.
- Thunderbolt 5 RDMA: Low-latency communication helps the machines exchange intermediate results efficiently.
How Fast Is It on Multiple Mac Studios?
Apple's benchmark tested Muse Spark Open using two and four M5 Ultra Mac Studio systems, with the model evaluated in NVFP4 precision. At an 8,192-token context size, Apple's published results show the following generation rates.
| Configuration | Reported Generation Speed | Important Detail |
|---|---|---|
| Two Mac Studios, pipeline parallelization | 15.74 tokens/sec | Model layers are split between two systems |
| Two Mac Studios, tensor parallelization | 24.99 tokens/sec | Computation is distributed across both systems |
| Four Mac Studios, tensor parallelization | 36.73 tokens/sec | More parallel compute, with additional hardware cost |
These are Apple's measurements under a defined benchmark, not a universal guarantee for every prompt, context length, or agent workflow. The results nevertheless demonstrate that distributed inference over Thunderbolt can make a very large open-weight model usable on a desktop cluster.
The Catch: This Is Not an Ordinary Mac Setup
The hardware requirements make this a niche demonstration for now. Meta's and Apple's examples use high-memory Mac Studio systems; The Wall Street Journal reported that the required hardware could cost more than $40,000 depending on the configuration.
That figure should not be interpreted as the price of the model itself. Muse Spark Open is available to download under Meta's Muse Model License, while the expensive part is assembling enough suitable memory and compute to run it locally at useful speeds.
Why run it locally?
- More control over where prompts and model inference run.
- Potentially lower recurring API costs for sustained use.
- Greater flexibility for customization and development.
- A useful test bed for private, local agent workflows.
What makes it difficult?
- Very high upfront hardware cost.
- Model setup and distributed configuration require technical work.
- Power use and inference speed depend on the cluster.
- Local execution does not automatically guarantee privacy or security.
Apple Mac Studio
For developers considering local model inference, the Mac Studio is the relevant hardware family. However, Muse Spark Open requires a high-memory configuration and distributed setup; a standard desktop configuration should not be assumed to run it by itself.
Check Mac Studio Listings on AmazonThunderbolt 5 Cables
Apple's distributed MLX setup uses two Thunderbolt cables between connected Mac Studio systems to increase link bandwidth. Compatible certified cables are necessary, but the cable alone does not provide distributed inference without the required software and hardware configuration.
Explore Thunderbolt 5 CablesIs Muse Spark Open Really Open Source?
It is more precise to call Muse Spark Open an open-weight model. Meta provides downloadable weights under its Muse Model License, which allows broad use and modification, including commercial uses subject to the accompanying usage policy.
That is useful for developers who want to inspect model behavior, run inference on their own infrastructure, or customize the model. But downloadable weights alone do not mean that all training data, training code, and every detail of the development process have been released.
Important privacy distinction
Running inference locally can keep prompts off a remote model API, provided the application is configured accordingly. Agents may still access websites, call external services, or transmit data through tools, so local execution is not a substitute for checking permissions and network behavior.
What This Means for AI Developers
Muse Spark Open expands the options for developers who want large, multimodal models with more control over deployment. Its open-weight distribution also lets developers access it through supported hosted inference providers, meaning a Mac Studio cluster is not the only way to experiment with the model.
The larger shift is that AI hardware is becoming a more flexible collection of choices: a cloud API for occasional use, a single workstation for smaller local models, or a connected cluster for enormous models that need more memory. The right choice depends on usage, latency, privacy requirements, and total cost.
“The future is for everyone, and that includes the freedom to run advanced models wherever.”
Meta's stated rationale for releasing Muse Spark Open. Source: Meta AI Research.
Final Take: A Meaningful Collaboration, With a Very High Entry Price
Apple and Meta have demonstrated that a frontier-scale open-weight model can run locally on a cluster of high-memory Mac Studios. The collaboration is notable not because every Mac can suddenly run an 816-billion-parameter model, but because MLX, quantization, unified memory, and fast interconnects make distributed desktop inference possible.
For most users, hosted inference or a smaller local model remains the more practical route. For AI researchers and developers who need to experiment with very large open-weight systems, this collaboration offers an intriguing new path.
Explore More AI and Computing Tools on Solidaitech
Planning your own local AI setup? Explore Solidaitech's technology tools and calculators to understand memory requirements before buying hardware. Our Apple Silicon AI RAM Calculator can help estimate which supported local language models fit your Mac's configuration.
Try the Apple Silicon AI RAM CalculatorSources and further reading
Meta AI Research — Introducing Muse Spark Open, October 15, 2026
Hugging Face — Muse Spark Open model card and download
Meta — Muse Model License and Usage Policy
CNBC — Apple and Meta partner on AI for Mac
The Wall Street Journal — Meta and Apple team up to put open-weight AI on Macs
Axios — Apple and Meta bury the hatchet for AI
John Ternus on X — Response to the Muse Spark Open collaboration
Frequently Asked Questions About Apple, Meta, and Muse Spark Open
What is Muse Spark Open?
Muse Spark Open is Meta's downloadable open-weight AI model with 816 billion total parameters, 42 billion active parameters per token, and a one-million-token context window. It is designed for coding, multimodal understanding, and complex agentic tasks.
Can Muse Spark Open run on a single Mac Studio?
Apple's technical report says the quantized model occupies about 493GB for weights alone, with additional memory required for inference computations. A single Mac Studio with up to 512GB of unified memory cannot accommodate the complete workload, so the demonstrated setup distributes it across multiple systems.
How fast does Muse Spark Open run on Mac Studio?
Apple measured 36.73 tokens per second using four M5 Ultra Mac Studio systems with tensor parallelization at an 8,192-token context size. Two-machine configurations were slower in the same reported test; results vary by context and workload.
Is Muse Spark Open free to download and use commercially?
The model weights are available to download under Meta's Muse Model License, which permits many uses, including commercial use, subject to the license and usage policy. Developers should review those terms before deploying or redistributing the model.
Does Apple's collaboration with Meta mean Siri will use Muse Spark Open?
No such product integration was announced. The October 2026 collaboration concerns optimizing Muse Spark Open for local inference on Mac Studio through Apple's MLX framework; a future Siri integration has not been established by the cited announcements.
No comments:
Post a Comment