The open-weight system, unveiled on Thursday, contains 816 billion parameters and can run locally across two Mac Studios equipped with Apple’s M5 Ultra processors. Its release allows developers to download the model’s underlying weights rather than depend exclusively on remotely hosted services.
Meta said the model was developed for complex software engineering, coding assistance and autonomous tasks requiring several stages of reasoning. It has 42 billion active parameters at a time and supports a context window of one million tokens, enabling it to process substantial amounts of information during a single interaction.
The collaboration brings together companies that have clashed over digital advertising and privacy policies. Meta chief executive Mark Zuckerberg acknowledged Apple’s contribution to the optimisation work, while Apple chief executive John Ternus welcomed the ability to operate a model of this scale on desktop hardware.
Meta’s artificial intelligence chief Alexandr Wang said the memory capacity of Mac Studio systems made it possible to run models that would otherwise exceed the capabilities of ordinary personal computers. Local execution also means developers can process prompts and proprietary information without automatically transmitting them to Meta’s servers.
The model uses a mixture-of-experts architecture, activating only part of its overall parameter count for each operation. That approach reduces the computing demands associated with generating responses, although the complete model still requires substantial memory to remain accessible.
Apple’s M5 Ultra Mac Studio can provide its graphics processor with up to 512 gigabytes of unified memory. Connecting two machines creates sufficient combined capacity to accommodate the model, using Apple’s MLX software framework to distribute processing between devices.
Apple’s technical measurements indicate that a four-machine configuration can generate more than 36 tokens per second, while a two-machine arrangement achieves approximately 25 tokens per second. Performance depends on the configuration and workload, and the hardware requirements put local deployment beyond the reach of many individual users.
Meta has also made the model accessible through hosted application programming interfaces for developers who do not own the necessary equipment. Downloadable weights are available through Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models.
The company said Muse Spark Open supports several inference systems, including vLLM, SGLang and llama. cpp, alongside Apple’s MLX. It can also be adapted using the PyTorch framework, allowing developers to tailor its behaviour for particular applications.
Meta demonstrated the model alongside its Muse Code programming tool. Tests included making changes to an open-source Mastodon application and controlling DaVinci Resolve video-editing software to identify and crop selected footage. These demonstrations illustrate intended uses but do not establish performance across every production environment.
Meta described Muse Spark Open as the most capable open-weight model developed in the United States, citing its own evaluations. Such benchmark claims depend on the tests selected and should not be treated as proof of superiority across all tasks or competing systems.
The weights are distributed under the Muse Model License, which permits many developers to use them without payment but imposes commercial restrictions. Open-weight availability therefore does not mean the software carries an unrestricted open-source licence.
Meta said the system underwent additional safety training and was assessed under its Superintelligence Scaling Framework before release. The company stated that it met the thresholds required by that framework, with a fuller evaluation report expected to provide additional detail.
The arrangement gives Meta another distribution route for its artificial intelligence technology while providing Apple with a demanding demonstration of its desktop processors. Neither company has announced that the collaboration extends to consumer assistants or other products, and the release does not establish a broader commercial partnership between them.
The ability to operate a model without sending every request to a remote service may appeal to software teams handling confidential code or sensitive internal documents. Nevertheless, local processing does not remove the need for developers to assess model accuracy, security and compliance when incorporating generated material into applications or business processes under their own controls.
The Mac deployment also requires connected machines rather than a single standard desktop. Apple’s unified memory design enables the processor and graphics hardware to share available capacity, an important consideration when loading models with hundreds of billions of parameters.
Follow Arabian Post
Select Arabian Post as your preferred source on Google and MSN News for trusted business news and Arab politics and updates.
