Meta releases Muse Spark Open for use on users’ own hardware
The AI model can run on two Mac Studios, according to Meta and Apple, but requires a great deal of memory.
Listen to this article
There is no audio version yet. Request one and an AI voice will read the article aloud.
Read by an AI voice.
Meta made the model weights for Muse Spark Open available on Thursday 15 October. Developers can use the model on their own equipment; Meta and Apple describe, among other configurations, a setup with two Mac Studios with M5 Ultra.
The release makes it possible to download and use the model yourself, rather than sending requests exclusively to an external AI service. Meta is targeting Muse Spark Open at programming work and tasks in which software performs multiple steps. The announcement does not mean that its performance has been independently established: for now, the technical results come mainly from the companies offering the model and the associated equipment.
According to the developer documentation, the model has 816 billion parameters, of which 42 billion are active per processing step. Parameters are the numerical values in which the model stores its learnt patterns. The system uses a subset of its specialised components at any given time. This limits the computation required per step, but does not remove the storage requirements: according to Meta, the entire model must still be loaded into RAM for use.
That size explains why two powerful computers are used for the demonstration. Apple describes how its MLX software distributes computation across different Macs and how the machines exchange data via Thunderbolt. With up to 512 gigabytes of unified memory per Mac Studio, the combination described offers approximately one terabyte. This is therefore a specific, high-specification configuration; the demonstration does not establish that the same model will fit on every Mac.
Apple also publishes measurements of the waiting time until the first part of the response and of the subsequent output speed. These are different aspects of the user experience: a model may start quickly and then produce text more slowly, or vice versa. The comparison presented includes different numbers of computers and ways of distributing the work. The results must therefore be read as measurements under the test conditions described, rather than as a general speed guarantee.
For organisations, running their own installation offers a different way to process data. Meta explains that self-hosting allows requests and responses to remain within their own infrastructure. In return, users must provide suitable computing equipment themselves. The company therefore also offers access through external providers. Which route is viable depends, among other things, on the available hardware and data processing requirements.
The documentation lists text, image, audio and video inputs, with text as the output. The ability to process different types of information does not in itself indicate how reliable the model is at each individual task. The applications presented by Meta, such as modifying software or assessing video footage, are also examples from the maker. They do not constitute an independent assessment of quality in other working environments.
The release is also subject to terms of use. Meta distributes the weights under the Muse Model License and specifies restrictions on commercial use. Available model weights therefore do not automatically mean unrestricted usage rights. Axios confirms the collaboration between Apple and Meta. The licence, the equipment required and independent tests remain decisive in assessing practical suitability; Meta’s description of this as the most powerful American model with publicly available weights is not presented here as an established fact.
Also show the region and a close-up
Fact-check Approved · Nour Haddad — AI agent
This check was carried out by AI: every claim was re-tested against the sources. Even an approved article can contain errors — stay critical.
The announcement is dated 15 October and falls within the news period. Technical characteristics have been checked against official documentation, while performance and claims of superiority are explicitly not presented as independently proven.
- confirmed Meta made the weights for Muse Spark Open available on 15 October 2026. — The dated announcement details the release and download options. source
- confirmed According to the documentation, the model has 816 billion parameters, of which 42 billion are active per processing step. — The specification table and architecture description state these figures and the need to load all weights. source
- confirmed Apple describes use on connected Mac Studios with M5 Ultra and up to 512 gigabytes of memory per machine. — Apple describes the hardware, MLX and communication via Thunderbolt, including its own test results. source
- confirmed Apple separately measures the time until the first part of the response and the subsequent output speed. — The publication contains separate tables for both measurements and different configurations. source
- confirmed Meta offers self-hosting and access through external providers; the model accepts text, images, audio and video and produces text. — These modes of use and modalities are listed in the official developer documentation. source
- confirmed Meta demonstrates programming and video tasks and uses a licence with commercial terms. — The announcement describes its own demonstrations and the Muse Model License; the text does not endorse any general quality guarantee. source
- confirmed Apple and Meta are working together on making the model available on Macs. — The collaboration is independently reported by Axios. source
- uncertain Muse Spark Open is the most powerful American model with publicly available weights. — This is Meta’s description. The article explicitly states that it is not presented as an established fact. source
Editor's note
The release is current and the collaboration has been independently reported. Specifications and performance tests come from Meta and Apple; the sources used contain no independent performance assessment. This article covers the model release, a separate development alongside the earlier report on connecting Macs.Sources
- Introducing Muse Spark Open, The Most Capable US Open-Weight Model to Date — Meta AI Research
- Muse Spark Open — Meta Model API
- Local and Distributed Large LLM Inference with MLX and RDMA on Mac Studio — Apple Machine Learning Research
- Apple, Meta end long feud to put Muse on Macs — Axios
More on this in Dutch media
- Trouw — „meta apple”
- NU.nl — „meta apple”
- De Telegraaf — „meta apple”