Apple and Meta have spent years getting in each other's way more than helping each other: fights over app tracking, clashing views on privacy, and when Apple needed a flagship model for Apple Intelligence, it partnered with OpenAI, not Meta. Against that backdrop, the October 15 release looks almost like a truce.
Meta unveiled Muse Spark Open, an open-weight model with 816 billion parameters, though only 42 billion are active at once thanks to a mixture-of-experts design, plus a one-million-token context window. The real trick isn't the specs — it's what the model was built to run on: two Mac Studios with M5 Ultra chips can pool their memory into a single terabyte and hold the entire model, acting as one machine.
"Open-weight" means the weights can be downloaded and run on your own hardware instead of accessed only through someone else's API. The Muse Model License lets most developers use it for free, with some commercial limits attached. The model is already up on Hugging Face, Amazon SageMaker, Google Cloud and Microsoft, plus services like Together AI and OpenRouter.
Meta's demos weren't built for chit-chat. Paired with its Muse Code terminal, the model fixed bugs in a Mastodon client and edited footage in DaVinci Resolve — work meant to show off multi-step, agentic tasks rather than simple Q&A.
It's Meta's second big model news in a week: days earlier, the company showed off a closed, top-tier model that beat nearly every rival except Claude. Spark Open is that release's open sibling, aimed at developers and small teams rather than casual users — a terabyte of memory means a loaded Mac Studio, not a laptop.
The real story isn't the benchmark scores. It's that Apple agreed, for the first time, to tune its own hardware around a rival's model instead of pushing only its in-house Apple Intelligence stack.


