Meta released Muse Spark Open on Thursday, Oct. 15, an open-weight AI model with 816 billion parameters and a 1 million token context window. The weights are on Hugging Face, and Meta tuned the model for the Mac Studio with Apple’s M5 Ultra chip. The catch: running the whole thing locally takes roughly 1TB of memory.
What does “open-weight” mean?
An open-weight model is one whose “weights,” the numbers training adjusts to store what the AI learned, are available to download. You can run it on your own hardware instead of relying on the company’s servers. That is not the same as fully open source. Training data and other details can stay private, and the reports we drew on do not say which license Meta attached to Muse Spark Open.
Meta calls it “the most capable US open-weight model to date.” By Meta’s own benchmarks, it beats two leading Chinese models, GLM 5.3 and Kimi K3, on agentic tasks (where the AI carries out multi-step work on its own) and visual reasoning. No detailed numbers or independent tests are out yet, so treat that as Meta’s claim. The company also said it believes the US should not let Chinese labs lead in open-source AI.
What can it do?
Muse Spark Open is natively multimodal, meaning it was built from the start to handle more than one kind of input, such as text and images. Of its 816 billion parameters, only 42 billion are active at any moment, a design that cuts the work done for each word generated. Meta aims it at coding, long-running tasks and multi-step reasoning.
To show real use, Meta ran workflows entirely on-device, pairing the model with Muse Code, its AI coding terminal. In one test it fixed bugs, added features and produced a full build of Ice Cubes, an open-source Mastodon client for iOS. In another it worked through DaVinci Resolve, the video editor, reviewing footage frame by frame, picking the best moments and cutting them into vertical clips.
Why the Mac Studio, and what does it cost?
The reason is the unified memory in Apple silicon, which the CPU and GPU share. According to Meta, the M5 Ultra can address up to 512GB per Mac Studio, and its 1.2TB/s of memory bandwidth is enough to read the small slice of the model needed for each token. Two 512GB Mac Studios linked together act like one supercomputer with 1TB of unified memory, in Meta’s description. Meta used the latest version of MLX, Apple’s open-source framework for its chips, and says the pairing is the best way to run a powerful model on your own desktop.
The price tag is steep. Apple has not announced pricing or a ship date for the 512GB Mac Studio, which is expected in late October. Engadget estimates that a setup of four 256GB Mac Studios would run about $38,000. In its view, that puts local use out of reach for individual developers, though it could make sense for large organizations that spend heavily on tokens, if the model performs as claimed.
Can you use it without that hardware?
Yes, through an API. TechCrunch reports the model is already available on several API platforms, without naming them, and you pay per use rather than buying the machines. Pricing on those platforms was not reported.
The Apple and Meta backstory
The partnership stands out given the two companies’ history of friction. In an interview with Axios, Meta AI chief Alexandr Wang said the companies have worked together for several months and that he is glad to see a warmer relationship. He credited Daniel Gross, a former Apple executive, with an important role in bringing them closer.
The release lands days after Meta Muse Wonder, the company’s first frontier model, which is closed and billed per use through the API. Muse Spark Open goes the other way: weights anyone can download. For background on why these systems sometimes get things wrong, see our explainer on what a language model is.
Is it worth caring about?
If you are an everyday user, the direct impact is small. Nobody is running 816 billion parameters on a laptop. The likelier effect is secondhand: developers and companies building services on top of it, and more competitive pressure on Chinese open models. If you are a developer, watch two things: the license, which you should check on the model page before any commercial use, and the final price of the 512GB Mac Studio. For occasional testing, an API provider will cost far less than building the rig.
