Meta launches open-weight Muse Spark Open model that can run on Apple Macs

What happened

Meta has launched Muse Spark Open, which it calls the “most capable” open-weight model in the United States. According to TechCrunch, Meta released evaluation results on Thursday claiming the model performs better than other popular open-weight models [1]. Open-weight means the model’s weights are published, which TechCrunch says can make models more cost-efficient and offer more transparency and flexibility than closed models [1].

The model is described as natively multimodal and is built for complex coding tasks and agentic work [1]. Meta says it “excels at long-horizon tasks, complex software engineering, and multi-step reasoning” [2]. Hugging Face describes it as a mixture-of-experts model that handles text, images, audio, and video, with about 816 billion parameters, around 42 billion active per token, and a context window of up to one million tokens [3].

The model is designed to run on Apple computers. Meta says it can run on desktop using two Apple M5 Ultra Mac Studios with the latest MLX framework, and calls it the first model in its class to do so [2]. TechCrunch says the best performance is expected on the Mac Studio with the M5 Ultra chip, but that running it locally still requires a lot of power, with at least 1 terabyte of RAM according to Engadget [1].

Hugging Face reported different hardware results. It tested quantized versions on a Mac M3 Ultra with 512GB of memory. A 2-bit quant ran at 18 tokens per second, with the model using 304 GB [3]. Hugging Face also said it tested Q2_K, Q4_K_S and a Q8 variant on Apple Silicon through llama.cpp [3].

Meta paired the model with Muse Code, its AI coding terminal. According to TechCrunch, Meta used it for coding fixes on a Mastodon client, including bug fixes, new features, and a full build. Meta also said the model could review and crop footage in DaVinci Resolve [1].

According to The Verge, Meta and Apple, which are typically rivals, worked together on the launch and publicly support each other. Meta CEO Mark Zuckerberg congratulated the team on X, and Apple CEO John Ternus replied that running a model of this scale so efficiently from a desk is impressive [2]. Meta chief AI officer Alexandr Wang told Axios the companies have collaborated for several months [2]. Apple also published a technical post about MLX, an array framework for machine learning on Apple Silicon, covering Muse Spark Open and two other open models [2].

Meta is making the model available now through a number of API platforms [1]. The sources do not state the cost of using these platforms.

Who is affected

Developers are the group the sources name most directly. Hugging Face says the model gives developers a common interface for asking questions about different kinds of input [3]. Mac owners with high-memory machines are also affected, since the sources describe local running on Mac Studio hardware [1][2].

Hugging Face also notes that reasoning cannot be turned off in Muse Spark Open. It advises keeping reasoning set to “low” when a task is easier, to use fewer tokens [3]. For deployment outside a Mac, Hugging Face lists an SGLang recipe for an NVFP4 variant on NVIDIA RTX Pro 6000 hardware, and says 8 NVIDIA H200 141GB GPUs are an option for the full 1,048,576-token context length [3].