Meta Muse Spark Open is large enough to sound like a model that belongs only inside a remote data centre. Yet on 15 October 2026, Meta released its weights and Apple published an engineering demonstration of the model running locally across Mac Studio computers. The development is a notable collaboration between companies that usually compete for control of consumer technology. It does not mean that an ordinary MacBook can now run a frontier-sized model at full quality. It means sufficiently equipped developers may have a new way to run powerful software on hardware they control, potentially changing the balance between cloud access and local experimentation.
Meta’s announcement of Muse Spark Open describes an open-weight model with 816 billion parameters in total, of which around 42 billion are active for a given token. It can work with text, images, audio and video and is designed for complex coding and agent tasks. Those numbers describe the model’s architecture, not a proven claim that its answers are better than every competitor’s. Meta calls it the most capable American open-weight model to date, which is the company’s own positioning. Independent evaluations will be needed to judge performance in varied real work.
What Meta Muse Spark Open needs to run locally
Large language models normally demand substantial memory, powerful processors and efficient software. In this case the distinctive element is Apple’s MLX framework, which is designed for machine learning on Apple silicon. Meta says the model can be run using two Mac Studios with M5 Ultra chips. Apple’s engineering write-up explores configurations including clusters of four machines, where it measured output exceeding 36 tokens per second under its stated conditions. That is a controlled technical result, not a speed promise for every task or every user.
Apple’s local inference research post explains how unified memory and fast connections between machines make a large model workable on desktop equipment. Its tests compare several models and distinguish initial response delay from the speed of generated text. The model’s design activates only part of its total parameters at a time, which reduces the processing required for each output step. Even so, distributing a model of this size requires serious hardware resources. Describing the system as ‘running on a Mac’ without explaining the multi-machine setup would give consumers the wrong impression.
There is a broader difference between open weights and a fully open service. Releasing model weights allows developers to download and run the model under its licence, but they still need suitable hardware, compatible software and responsibility for deployment. The model is also offered through cloud platforms, which may be easier for people who do not want to operate their own equipment. Local operation can provide more control over data flow and testing, but only if the surrounding application is also configured to keep the relevant information local.
Why rivals would work together on this
Meta gains more ways for developers to use its model. Apple gains a demonstration of what its high-end hardware and machine-learning software can handle. Their incentives align even if the two companies remain competitors elsewhere. It is a reminder that the AI industry is not organised into neat permanent camps. A company can build a model, another can make the hardware efficient for it and developers can choose how to deploy the result. That does not remove commercial rivalry, but it may widen the selection of tools available outside a single cloud provider.
Meta describes demonstrations involving coding workflows and multimedia software. As with other vendor examples, these show selected uses rather than established success rates across all possible applications. A local model can still make mistakes in code, misinterpret footage or take an incorrect action if connected to tools with broad permissions. Running it on equipment in an office does not make it automatically trustworthy or safe. Users still need checks, access controls, testing and a plan for recovering when the system behaves unexpectedly.
LiveAIWire recently covered Meta’s Muse Wonder models and longer-task ambitions. Muse Spark Open is a distinct announcement because it concerns open weights and alternative deployment, not simply another account of the same product’s memory abilities. It also intersects with small-scale local AI hardware development. The central question is who controls the machine that performs the work and what compromises follow from that choice.
Local control is useful, but it is not free
Cloud models spare users the expense and maintenance of a complex computing setup. Local installations may appeal to organisations that have confidential material, specialised applications or predictable ongoing workloads, but the financial and operational trade-offs depend on hardware costs, electricity, staff expertise and usage patterns. Nobody should infer that two or four high-end desktops are cheaper than a cloud service for every company. The Apple and Meta announcements establish technical feasibility, not a universal cost advantage.
The significance of the release is therefore broader than a performance chart. It demonstrates that the route to powerful AI need not always pass through the same remote provider interface. Developers can evaluate a large model, decide where to operate it and explore local applications with a different set of data-handling choices. For most consumers this remains a technical story rather than a new button on their computer. For the businesses building future tools, it is a meaningful expansion of where the intelligence can live.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.