Two of Silicon Valley’s most famous rivals just teamed up. On Thursday, Meta released Muse Spark Open, a massive new open-weight artificial intelligence model, in direct partnership with Apple. The headline claim: this 816-billion-parameter model, built for coding and autonomous AI agents, can run entirely on a desk full of Mac Studios instead of inside a cloud data center.
It sounds like press-release fantasy until you look at the engineering. Two linked Apple Mac Studios, each carrying the new M5 Ultra chip, can now hold and run one of the most capable open AI models on the planet.
What Muse Spark Open actually is
Muse Spark Open is what Meta calls its most capable open-weight model developed in the United States so far. The numbers are serious: 816 billion total parameters, with 42 billion active at any one time thanks to a mixture-of-experts design that only wakes the parts of the model each task needs.
The model supports a context window of up to one million tokens, so it can read and reason over enormous documents or codebases in a single pass. It is natively multimodal, handling text, images, video and audio, and Meta has tuned it for coding, complex software engineering, multi-step reasoning and long-running agentic tasks, the kind of work where an AI plans, acts, checks its own output and keeps going.
On benchmarks shared by Meta, it matches or outperforms other leading open-weight systems such as Kimi K3 and GLM 5.3 on certain tests.
How Apple fits into the picture
The surprise is not just the model, it is who Meta built the local-computing story with. Meta chief executive Mark Zuckerberg announced the release on X and credited Apple chief executive John Ternus and his team; Ternus replied with congratulations and pointed to the Mac Studio’s on-device AI credentials.
The recipe is clever. Even compressed to 4-bit precision, the model’s weights occupy roughly 493GB. Two Mac Studios, each built around the M5 Ultra chip with 512GB of unified memory and 1.2TB/s of memory bandwidth, together provide about a terabyte, enough for the whole model.
Apple’s open-source MLX framework splits the model across the machines, which talk over Thunderbolt 5 using remote direct memory access, a technique that moves data between computers with very low latency. Apple published its own benchmarks: a two-Mac setup generated about 25 tokens per second, and a four-Mac configuration hit nearly 37. Usable, not blazing, but this is a frontier-scale model running on desks, not in a server hall.
What Meta showed it doing
To prove the point, Meta demonstrated its Muse Code agent running the model locally. In one demo, the agent fixed bugs, added features and completed a full build for Ice Cubes, an open-source Mastodon app. In another, it took control of DaVinci Resolve, watched through video footage, picked the best moments and cropped them into vertical clips.
Those are the long-horizon, multi-step workflows Meta keeps talking about: not answering a single question, but carrying a project across dozens of steps. Meta’s AI chief Alexandr Wang said the memory on Mac Studios is what makes a model of this size practical to run.
Why open weights change the maths
Open-weight models publish their trained parameters so developers can download, modify and deploy them, unlike closed models that stay under their creator’s control. Since processing happens on local machines, Meta itself never sees user prompts or usage data.
The model is available through API providers including Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI, and weights can be downloaded from Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models.
The catches: price, license and safety
None of this comes cheap. A single 256GB Mac Studio starts at $10,799, and the 512GB configuration for the two-machine setup only ships later this month. A four-machine configuration can run past $40,000. Frontier-scale AI now fits on a desk; it just costs about as much as a car to get there.
The weights are released under Meta’s Muse Model License, free for many developers but under limited commercial terms, and Meta says the model met the thresholds of its internal safety framework.
For India’s large developer and startup community, the interesting part is the direction of travel. Open-weight frontier models that run outside the cloud cut API bills and keep sensitive code on local hardware, a meaningful argument for startups and labs on tight budgets.
FAQs
What is Meta’s Muse Spark Open?
Meta’s new open-weight AI model with 816 billion total parameters, built for coding and autonomous AI agent tasks, released on October 15, 2026.
How can it run on Apple Mac Studios?
Two to four M5 Ultra Mac Studios are linked over Thunderbolt 5 using Apple’s MLX framework, pooling up to a terabyte of unified memory to hold the model.
Is Muse Spark Open free to use?
Weights are downloadable and free for many developers under Meta’s Muse Model License, though the license carries limited commercial terms.
Why is the Apple partnership significant?
Apple and Meta have historically kept their distance, so a direct engineering collaboration on locally run open AI marks a notable industry shift.
Meta and Apple keep pushing the frontier in opposite directions, one open, one premium. This week those directions crossed. We will be watching what developers build with the combination.
