Meta releases Muse Spark Open, an open-weight AI model designed to run on Macs
According to TechCrunch, the 816-billion-parameter model is built for coding and agent tasks but needs at least 1 terabyte of RAM to run locally.
Meta has released a new open-weight AI model called Muse Spark Open and says it is designed to run locally on Apple's Mac computers, according to TechCrunch.
Meta calls the model the "most capable" open-weight model in the United States, TechCrunch reported. The model has 816 billion parameters and a context window of 1 million tokens. Meta says it outperforms other popular open-weight models, based on evaluation results the company released on Thursday.
According to TechCrunch, Muse Spark Open is natively multimodal and built for complex coding tasks and agentic work, meaning tasks the model carries out by itself. Running it at home will take serious hardware. TechCrunch, citing Engadget, reported that the model needs at least 1 terabyte of RAM. The report says it runs best on a Mac Studio with the M5 Ultra chip.
Meta has also shown what it says the model can do. The company paired Muse Spark Open with Muse Code, its AI coding terminal, and used it in several agent-style workflows, TechCrunch reported. One was work on a Mastodon client that Meta said included "bug fixes, new features, [and] full build." Meta also says it used the model to run DaVinci Resolve, the video editing software, where it reviewed and cropped footage.
The release matters because of the kind of model it is. TechCrunch notes that open-weight models can cost much less to use than closed, proprietary ones, and can offer more transparency and flexibility. A model this large that can run on a single desktop machine, even an expensive one, could let developers do demanding coding work without sending it to a cloud service.
Meta is making the model available now through several API platforms, according to TechCrunch.
Some questions remain open. The performance claims come from Meta's own evaluation results, and the report does not mention any independent testing. TechCrunch does not say which competing models Meta compared against, which API platforms carry the model, or what it costs to use. The report also does not describe the licence terms for the model's weights. The hardware requirement comes from Engadget as cited by TechCrunch, and this piece rests on that single TechCrunch report.
