The model is designed for software development, multi-step reasoning and longer-running agent tasks. It is also natively multimodal, allowing it to work with text, images, audio and video as inputs.
The Apple configuration is a central part of the release. Meta says two M5 Ultra Mac Studios can hold the full model using Apple’s unified memory architecture, with each machine supporting as much as 512GB of memory. CNBC reported that customers can link between two and four Mac Studios depending on their setup.
Running Muse Spark Open locally gives developers an alternative to sending their workloads to cloud data centers. Apple and Meta say that can offer additional security while reducing dependence on hosted computing services.
Meta AI chief Alexandr Wang said the size of many advanced open-weight models makes local deployment impractical on ordinary personal computers, while the memory available in Apple’s Mac Studio makes running Muse Spark Open possible.
“We worked with Apple to bring leading Agent capabilities to your desktop for the first time on the Apple M5 Ultra Mac Studios, so you know we’re really excited about this,” Wang said.
Meta demonstrated the model performing agentic tasks entirely on-device. One example used Muse Code, powered by Muse Spark Open, to make changes to the open-source Ice Cubes Mastodon client, including fixes and new features. Another had the model operate DaVinci Resolve, review video footage and prepare vertically cropped clips.
Developers downloading the model for Apple hardware can use MLX, Apple’s machine-learning framework. Meta says Muse Spark Open has been tuned specifically for Apple silicon, while Apple has published additional technical documentation showing how MLX can distribute large open models across multiple machines.
Muse Spark Open also supports a wider set of deployment options. Meta lists SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, llama.cpp and Unsloth among the compatible inference tools, with support spanning hardware from AMD, Apple, Intel and Nvidia. Developers can further customize the model through PyTorch.
The hardware requirements are substantial. CNBC reported that a Mac Studio with an M5 Ultra chip and 256GB of memory starts at $10,799, while a version offering 512GB is scheduled for release later this month. Running the model locally requires more than one machine.
Meta is releasing Muse Spark Open under its Muse Model License, which permits many developers to use the weights for free while imposing some commercial restrictions. The model is also available through hosted services including Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI.
Weights can be obtained through Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face and Microsoft Foundry Models.
Meta says the model went through additional safety training before release. Its evaluations were conducted under the company’s Meta Superintelligence Scaling Framework, and Meta says Muse Spark Open met the thresholds required for an open-weight release.
The partnership also stands out because Apple and Meta have spent years at odds over privacy, advertising and control of computing platforms. Their relationship deteriorated further after Apple introduced privacy changes in 2021 that affected digital advertising companies, including Meta.
Apple’s leadership has since changed, with John Ternus succeeding Tim Cook as CEO last month. Meta CEO Mark Zuckerberg highlighted the new collaboration in a post showing Muse Spark Open running on a Mac.
“Muse Spark Open -- weights releasing today -- runs great on new Mac Studios thanks to work with @johnternus and Apple,” Zuckerberg wrote.
Ternus responded: “Congrats to the team! Amazing to run a model of Muse Spark Open’s caliber and scale this efficiently, right from your desk. Mac Studio keeps pushing the boundaries of on-device AI!”
Neither company said whether the work on Muse Spark Open would develop into a broader partnership.
The collaboration comes as Meta continues rebuilding its AI model lineup under Wang. The company previously centered much of its open-model strategy on Llama, but has since introduced several products under the Muse name.
Muse Glimmer, released in August, is a smaller 30-billion-parameter open-weight model designed for local agents. Meta has also released Muse Wonder, a more capable proprietary foundation model, while Muse Spark Open occupies a different position by giving developers downloadable access to a much larger model.
“These aren’t the most capable models that we’ve developed or released,” Wang said of Muse Spark Open. “But we think that there will be a lot of companies and developers who sort of use this model and tinker with it at small scale and then seek to ramp up and scale up their usage.”
Apple, meanwhile, has worked with several major AI model developers. It partnered with OpenAI in 2024 to integrate ChatGPT with Siri and announced an agreement with Google earlier this year involving Gemini models and Google Cloud infrastructure.
The Muse Spark Open collaboration adds local open-weight AI to that mix. Rather than integrating Meta’s model directly into an Apple consumer service, the companies are giving developers a way to run a very large AI model on Apple desktop hardware using downloadable weights and Apple’s own MLX software.
This analysis is based on reporting from Meta & CNBC.
Image courtesy of Meta.
This article was generated with AI assistance and reviewed for accuracy and quality.
