Meta brings an 816-billion-parameter open-weight model to a pair of Macs

Two Mac Studios can now host the same very large model. Meta released Muse Spark Open on 15 October, with downloadable weights and support for a pair of M5 Ultra machines running Apple’s MLX software. Developers can also use hosted providers, including Fireworks AI, Groq and Together AI, without assembling that desktop system.[1]

Muse Spark Open accepts text, images, video and audio and produces text. Its mixture-of-experts architecture contains 816 billion parameters, with 42 billion active during processing, and its stated inference context window is one million tokens. Those specifications describe the model; the desktop arrangement depends on a particular compressed representation and additional working memory.[3]

The connection between the computers matters

Apple puts the NVFP4 model’s memory footprint at about 493 GB. That sounds close to a Mac Studio’s maximum 512 GB, but intermediate computations also need room. Splitting the model across machines solves the capacity problem. In the simplest arrangement, one computer processes the earlier layers and passes its result to the computer holding the later layers.[2]

There is a faster division of labour: split work within each layer so the computers calculate together. That requires frequent exchanges, making the connection part of the inference system. MLX’s JACCL backend uses low-latency RDMA communication over Thunderbolt 5; its documentation requires direct connections between every pair of participating Macs.[2][5]

Apple measured 24.99 generated tokens per second on two M5 Ultra Mac Studios using this tensor-parallel arrangement, and 36.73 on four. The test used an 8,192-token context followed by 128 generated tokens. It gives a concrete indication of interactive speed at that workload, rather than the performance of the model’s maximum context window.[2]

This is a configured local cluster. The MLX instructions cover enabling RDMA in macOS Recovery, establishing the machine connections and describing them in a host file. The physical desktop size should not be confused with a one-click consumer installation.[5]

From model responses to a working coding tool

Meta reports using the local model through Muse Code to make and build changes to Ice Cubes, an open-source iOS Mastodon client. It also describes a DaVinci Resolve workflow that examined footage and cropped selected material vertically. These are Meta’s demonstrations, showing the sort of application connected to the model.[1]

Consider an illustrative development team working on an internal application. The model can examine an error report and propose an edit; the coding application opens the repository, applies the change and invokes the build. Keeping inference on the team’s hardware changes where the model request is processed. The rest of the workflow still depends on the software and services that the coding application calls.

A licence aimed differently at users and model hosts

The Muse Model License permits copying, modification, fine-tuning and distribution subject to its terms and usage policy. Products using the model or derivatives must display “Built with Muse Spark Open” in an allowed location, such as the interface, documentation, website or settings.[4]

The commercial distinction is substantial. A company can build around the weights without automatically becoming a model-hosting business; a provider selling access to models faces a separate condition as it grows. Alongside the Mac deployment route, that leaves two different opportunities: put a large model to work inside an application, or operate the service through which other developers reach it.[4]

Report an error or suggest a correction