Meta, the U.S. company, announced its open-weight model Muse Spark Open on Thursday local time. It has 816 billion parameters and a context window of 1 million tokens, and was designed with local operation on Apple computers in mind. It is already available through several API platforms.
Runs on a Mac—but not necessarily the Mac on your desk
When you hear “AI that runs on a Mac,” it’s easy to picture downloading it to your current laptop and giving it a try. But this is not quite that simple. Running a model of this scale—816 billion parameters—locally reportedly requires at least 1TB of memory. That’s an order of magnitude beyond the memory in a typical laptop.
The suggested optimal setup is a Mac Studio with an M5 Ultra chip. Rather than convenience, what Meta has demonstrated with this model is better understood as the option to handle a huge model on a device sitting on your desk instead of entrusting it to the cloud.
Still, the fact that this option exists is significant. Until now, ultra-large models have been on the other side of a data center. Now they are beginning to appear within reach of desktop computers. The requirements are steep, but the direction of travel is hard to ignore.
From AI that writes code to AI that works with video
Muse Spark Open has a natively multimodal design that can handle images and other inputs, and is built for complex coding tasks and agentic work. Agentic use means the AI does more than answer a question once: it plans multiple steps and carries out the work.
Meta says it paired the model with its AI coding terminal, Muse Code, and put it to work on various agentic workflows. One example was making changes to a client for the decentralized social network Mastodon. The company says the model handled the process from fixing bugs and adding features through to building the app.
The model was also used to operate the video-editing software DaVinci Resolve, where it could review and crop footage, Meta says. From coding to professional video-editing tools, the scenarios in which AI operates apps on screen are clearly expanding.
What I find particularly interesting is that both examples involve work on a desktop. AI is moving from answering in a chat window to acting like a colleague working at a PC. That may also help explain the emphasis on running the model locally on a Mac.
The value of open weights isn’t measured by performance alone
Based on the evaluation results it released, Meta claims the model outperforms other leading open-weight models. It describes Muse Spark Open as the highest-performing open-weight model in the United States.
But the significance of open weights cannot be captured by a leaderboard alone. Because the weights—the contents of the trained model itself—are released, open-weight models are said to have a cost advantage over closed models and to offer greater transparency and flexibility in how they can be used.
For businesses, the choice is whether to use the model conveniently through an API or run it in their own environment. The former is quicker to deploy; the latter keeps data from leaving the organization. The 1TB requirement is a high bar, but some teams handling sensitive development work or video footage may decide the investment is worthwhile.
Muse Spark Open gives organizations one more factor to consider. Try it through an API, then bring it in-house if needed. This two-stage approach may become a standard way to work with open-weight AI.
Comments