📰 Key Takeaways

Meta has launched what it calls its “most powerful open-weight model” in the US, Muse Spark Open, specifically designed to pair with Macs. The new model has 816 billion parameters and a 1 million token context window, and evaluation results Meta published Thursday show it outperforming other leading open-weight models. Muse Spark Open has native multimodal support and is mainly built for complex coding tasks and agentic workflows, while also being designed to run locally on Apple computers. But running the model locally isn’t cheap — according to Engadget, you’ll need at least 1TB of memory, and a Mac Studio with an M5 Ultra chip is the best setup to run it on. If you’ve got the hardware, Muse Spark Open could become a solid tool for coding and desktop work. Meta says it has paired the model with its own AI coding terminal, Muse Code, and deployed it across several agentic workflows — including “fixing bugs, adding features, and doing full builds” for a Mastodon client. Meta has also applied it to the professional video editing software DaVinci Resolve, showing the model can view and trim video clips. Compared to closed proprietary models, open-weight models offer advantages in cost-effectiveness, transparency, and flexibility. Meta has already made the new model available through multiple API platforms.

💬 JudyAI Lab’s Take

Meta has launched Muse Spark Open in the US, calling it the most powerful open-weight model out there right now — and the interesting part is that it’s specifically designed to pair with Macs. That choice of direction is worth paying attention to on its own.

For the AI builder community, this signals a shift: open-weight models are no longer just competing on parameter count — they’re starting to design toward specific hardware ecosystems, tying model capability directly to the end device. With 816 billion parameters and a 1 million token context window, built mainly for complex coding and agentic workflows, Meta has plugged it into its own coding terminal Muse Code, and even applied it to video editing software like DaVinci Resolve. That shows vendors are starting to treat open models as part of a complete toolchain, not just a standalone API product. But running it locally takes at least 1TB of memory, with a Mac Studio packing an M5 Ultra chip as the ideal setup — which tells you “open” doesn’t mean “accessible to everyone.” Hardware and software integration is the real barrier here.

If you’re trying to figure out whether this model fits your workflow, start by checking your hardware and actual task requirements before comparing parameter counts and benchmark scores.

📅 Source Information

🔗 Further Reading