Meta Muse Spark Open: an 816B open model that runs on two Mac Studios

Meta released the weights of a very large open model on October 15, 2026, and the most surprising part of the launch is not the model itself. It is the partner. Apple worked with Meta for months so that Muse Spark Open, an 816 billion parameter model, runs locally on a desk with two Mac Studio computers. For two companies that spent years attacking each other over privacy and ads, that is a remarkable photo op.

Here is what Meta actually shipped, what Apple measured on its own hardware, what it costs to try this at home, and why the move matters for the open model race.

The three things you need to know

• The model. Muse Spark Open is a natively multimodal mixture of experts model with 816 billion total parameters, of which 42 billion are active per token, and a context window of one million tokens. Meta built it for coding and agentic work and calls it the most capable US open weight model to date. The weights are on Hugging Face and several cloud model catalogs under the Muse Model License, which has limited commercial terms.
• The Macs. Meta says it is the first model in its class that runs on a desktop setup of two Apple Mac Studios with the M5 Ultra chip, using the newest version of Apple's MLX framework. Apple's own benchmark shows 15.74 tokens per second with two Macs in pipeline mode, 24.99 tokens per second with two Macs in tensor parallel mode and 36.73 tokens per second with four Macs, all at a context of 8,192 tokens.
• The alliance. Mark Zuckerberg publicly thanked new Apple CEO John Ternus for the collaboration. The two companies had not struck a public partnership since 2017, according to CNBC. The hardware is not cheap: a Mac Studio with M5 Ultra and 256 GB of memory starts at 10,799 dollars, and a 512 GB version arrives later this month.

What Meta released

Muse Spark Open belongs to the Muse family that Meta's superintelligence group has been building since it moved away from the older Llama line. The proprietary flagship, Muse Wonder, arrived only a few days ago, on October 12, and we covered it here on djcroman. The open model is a separate release, and Meta is clear that it is not its best system. Meta AI chief Alexandr Wang told CNBC that these are not the most capable models the company has built, but that he expects many developers and companies to tinker with it at small scale and then scale up their usage.

On paper the spec sheet is still huge. The model has 816 billion parameters in total, but because it is a mixture of experts design, only 42 billion of them do work for any single token. That is what makes local inference realistic at all: the machine has to store the whole model, but each step only needs to read a small slice of it. The context window is one million tokens, and Meta lists long horizon tasks, complex software engineering and multi step reasoning as the core strengths.

Distribution is broad from day one. Meta says the model is available through hosted API platforms at Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. The weights can be downloaded from Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models. Supported inference engines include SGLang, vLLM, TensorRT LLM, Hugging Face Transformers, Apple MLX, llama.cpp and Unsloth, with optimizations for AMD, Apple, Intel and NVIDIA chips. For fine tuning, Meta points developers to PyTorch and its TorchTitan library.

Meta also shipped custom GPU kernels called Helion inside SGLang. According to the company, they deliver kernel level speedups between 1.03 and 3.95 times on NVIDIA and AMD GPUs and a 3 percent end to end speedup on AMD's MI350.

The license is open, but not unlimited

Open weight does not mean open source in the strict sense. Meta releases Muse Spark Open under the Muse Model License with limited commercial terms. Meta says those terms allow many developers to use the model for free. If you plan to build a business on top of it, read the license text in the model card before you commit, because "limited commercial use" is exactly the kind of phrase that matters at scale.

How Apple makes 816 billion parameters fit on a desk

The Apple side of the story is the technical heart of the launch. Apple's machine learning research team published a detailed post the same day, together with a new MLX release.

The memory math is simple. In its native BF16 precision, Muse Spark Open would need more than 1.6 TB of memory just for its parameters. Apple tested it in a 4 bit format called NVFP4, which cuts the footprint to about 493 GB. Add the working memory needed during inference and it no longer fits into one machine, even one with 512 GB of unified memory. So Apple splits the model across several Mac Studios.

Meta describes the result in plain terms: the M5 Ultra's integrated GPU can use up to 512 GB of unified memory per Mac Studio, so two machines can hold the entire model, and the chip's 1.2 TB per second of memory bandwidth is fast enough to read the slice each token needs. In Meta's words, two Mac Studios behave like one supercomputer with a terabyte of unified memory.

Apple uses two ways to split the work. With pipeline parallelism, the first Mac runs the first half of the layers and hands the result to the second Mac. With tensor parallelism, every Mac holds a part of every layer, which is faster but needs many tiny messages between the machines. That only works because macOS 26.2 added a low latency RDMA driver over Thunderbolt 5, and MLX version 0.30.1 and later builds new communication tools on top of it. In Apple's setup, each pair of Macs is linked by two Thunderbolt cables to double the bandwidth.

The real numbers

Apple measured decode speed, which is how fast the model writes once it has read the prompt, with a context of 8,192 tokens:

• Two Mac Studios with pipeline parallelism: 15.74 tokens per second.
• Two Mac Studios with tensor parallelism: 24.99 tokens per second.
• Four Mac Studios with tensor parallelism: 36.73 tokens per second.

Apple describes the four machine result as well above average human reading speed. For very long prompts the cables matter too: with four M5 Ultras and a 65,536 token context, two cables per link cut the time to first token from about 50.5 seconds to 40.6 seconds in Apple's test.

The same post compares the new M5 Ultra with the previous M3 Ultra on smaller models that fit on one machine. Memory bandwidth rises from 819 GB per second to 1.2 TB per second, which Apple says translates into about 1.6 times faster token generation. Time to first token improves by up to 4.6 times thanks to the GPU Neural Accelerators in M5. Peak power during those tests was about 300 watts for a Mac Studio with M5 Ultra.

These are vendor benchmarks on vendor hardware, measured with Apple's own benchmark tool, so independent tests will matter. Still, they are concrete numbers with a published method, which is more than most local AI claims come with.

Meta's own agent demos

Meta says it tested several agentic workflows running entirely on device, using its Muse Code agent powered by Muse Spark Open. One demo shipped real iOS changes to Ice Cubes, an open source Mastodon client, including bug fixes, new features and a full build. Another drove the video editor DaVinci Resolve end to end, with the model reviewing footage frame by frame to find the strongest moments and crop them vertically.

Apple's post shows the practical route for developers. MLX can start a local server that speaks an OpenAI compatible API, and coding agents such as OpenCode can be pointed at that local address. In other words, the agent loop that usually goes to a cloud API can stay inside your own network.

What it costs to try this at home

This is where the "on your desk" pitch meets reality. CNBC reports that a Mac Studio with the required M5 Ultra chip and 256 GB of memory currently costs at least 10,799 dollars. A version with 512 GB of memory, the configuration Meta describes for the two machine setup, will be released later this month. Customers can pair between two and four Mac Studios, so a serious local setup quickly becomes a five figure purchase, before you count Thunderbolt cables, power and your time.

For a solo hobbyist that is out of reach. For a small company that pays a lot for cloud tokens, handles sensitive code, or simply wants an AI agent that never sends data off site, the math looks different. Meta and Apple both pitch the local route as more secure and cheaper over time.

Safety work on an open model

Open weights cannot be recalled once they are downloaded, and that raises the stakes for safety testing. Meta says Muse Spark Open received additional safety post training for that reason. The company evaluated it under its Meta Superintelligence Scaling Framework for the most severe risks and says it met the thresholds the framework requires for release. Detailed results are in the model card, and Meta says a full evaluation report is coming soon. Until that report is out, outside researchers have limited material to check those claims.

Why Apple and Meta suddenly work together

The relationship between the two companies has been cold for years. CNBC recalls that Apple's 2021 privacy changes limited ad targeting, and that Meta estimated in early 2022 that they would cut its revenue that year by 10 billion dollars. Tim Cook criticized Meta over privacy, and Zuckerberg said publicly that Apple had not invented anything great since the iPhone.

The tone changed with new leadership. John Ternus succeeded Cook as Apple CEO last month. On launch day Zuckerberg wrote that Muse Spark Open runs great on new Mac Studios thanks to work with Ternus and Apple, and Ternus congratulated the team. Wang credited Daniel Gross, a former Apple director who now works on his team, as one reason the collaboration happened. Neither company would say whether this is the start of a deeper partnership.

Apple already works with several AI labs. It partnered with OpenAI in 2024 to bring ChatGPT into Siri, later sued OpenAI in July over alleged trade secret theft, which OpenAI denies, and earlier this year struck a deal to use Google's Gemini models in Apple Intelligence. Running a top open model well on Mac hardware gives Apple another angle: if Nvidia owns the data center, Apple can try to own the high end desk.

Why it matters

• Local frontier class AI is now a product pitch, not a hack. Running models of this size used to mean a rack of GPUs. A launch that ships tuned weights, a framework update and published benchmarks on the same day turns local inference into something a team can plan around.
• The US open weight race heats up. Policymakers worry that free Chinese models could undercut American AI, as CNBC notes. Meta is positioning Muse Spark Open as the strongest US answer, while keeping its very best model, Muse Wonder, closed.
• Privacy and cost arguments get real numbers. With 24.99 tokens per second on two machines and 36.73 on four, coding agents can run at usable speed without sending source code to a cloud provider.
• Apple gets an AI story it can sell. Apple's own Siri update has drawn criticism, but showing an 816 billion parameter model running on its desktops is a strong hardware demo.

Dany's take

I love this launch for one reason: it proves that the line between cloud AI and personal AI is moving fast. Two years ago a model of this size was pure data center territory. Today it runs on two boxes that fit next to a monitor. That said, I would not order two Mac Studios tonight. More than twenty thousand dollars for a model that its own maker says is not its best only makes sense if you have a clear use case, like private code agents or sensitive documents. For most people the hosted APIs will be cheaper for a long time. The bigger signal is strategic: when Apple and Meta shake hands, the open model fight is no longer only about benchmarks, it is about who owns the hardware your agents live on.

What to watch next

• Independent benchmarks of Muse Spark Open on coding and agent tasks, beyond Meta's own evaluation report.
• The full safety evaluation report Meta promised, and how researchers react to it.
• Real world speed reports from developers once the 512 GB Mac Studio ships later this month.
• Whether Apple and Meta announce more joint work, or whether this stays a one time collaboration.
• How rival open model makers respond with their own local first releases.

Sources

• Meta AI Research: Introducing Muse Spark Open (15 Oct 2026)
• Apple Machine Learning Research: Local and Distributed Large LLM Inference with MLX and RDMA on Mac Studio (15 Oct 2026)
• CNBC: Apple and Meta team up to make Muse AI model run on Mac computers (15 Oct 2026)