By the Smili Phone team
On Thursday, October 15, Meta did something unusual for a big AI lab. It released the weights of Muse Spark Open, a giant AI model it calls the most capable US open-weight model to date. Anyone can download the model and run it outside of Meta's own apps and servers. Here is what that means, what it does not mean, and whether it matters for regular people.
What "open-weight" actually means
Most AI models you use, like the chatbots on your phone, run on the company's servers. You send a question, they send back an answer, and the trained model itself stays locked away. An open-weight model flips part of that arrangement. The trained weights, the huge file of numbers that makes the model behave the way it does, are free to download.
That gives developers real freedom. A company can host the model on its own computers, which means its data never has to leave the building. Developers can adapt the model for their own tools. Researchers can study how it behaves more directly. It is not the same as fully open-source software, though. Meta is releasing Muse Spark Open under its own Muse Model License, which allows limited commercial use, and it is not publishing everything that went into building the model. Still, it is a notable return to big open releases: in August, Meta released a smaller open-weight model called Muse Glimmer.
The specs that matter
Muse Spark Open is an 816-billion-parameter mixture-of-experts model, with 42 billion parameters active for any given token. In plain terms, it is a huge model that only switches on a fraction of itself for each word, which keeps it efficient. It accepts up to one million tokens of context, roughly a whole software project at once, and it is natively multimodal: it reads text, images, video and audio, and writes text back.
Meta built it for coding and agentic work: long multi-step tasks, tool use, and real software engineering. In Meta's own demos, the model shipped bug fixes and features to an open-source iPhone app and edited video footage end to end.
The catch: you need serious hardware
Downloading the model is free. Running it is another story. The full-precision version is over a terabyte, and even the compressed version is around half a terabyte. Meta worked with Apple to optimize it for the MLX framework, and the reference setup is two Apple Mac Studio desktops with M5 Ultra chips linked together, each with half a terabyte of unified memory. The snag: that 512GB configuration is not shipping yet. It is expected later this month. So no, you will not be running an 816-billion-parameter model on the laptop you already own.
How regular people can try it
You do not need a supercomputer to experience it. The weights are available on Hugging Face and the big cloud model catalogs, and seven hosting providers, including Groq, Together AI, Fireworks AI, OpenRouter, Baseten, Modal and Nebius, offer access from day one. That means you can experiment with the model through an API without downloading anything.
The bigger trend matters more than any single model. Each open-weight release pushes capable AI closer to running on hardware people actually own. A few years ago this would have been unthinkable. We tune up a lot of Macs here in Escondido, and the machines that will run tomorrow's AI well are the ones that are kept cool, clean and healthy today.
Related: running AI on your own Mac · OpenAI's DevDay and the agent wave
As an Amazon Associate, Smili Phone Repair earns from qualifying purchases.
