Meta has released its so-called "most powerful" open-weight model in the U.S., named Muse Spark Open. The model also features a key design: it is specifically optimized for Mac.
Meta's evaluation, released on Thursday, shows that this new model uses 816 billion parameters and a 1 million token context window, claiming superior performance compared to several other popular open-weight models.
Muse Spark Open natively supports multimodality and is designed for complex coding tasks and agent-based workloads. It is also optimized to run locally on Apple computers. However, running this model locally still requires substantial hardware; Engadget reports that at least 1 TB of memory is needed. The model performs best on a Mac Studio equipped with the M5 Ultra chip.
Meta stated that if sufficient memory is available, Muse Spark Open could become a key "workhorse" for coding and other desktop tasks. The company said it has integrated the model with its AI coding terminal, Muse Code, and deployed it across various agent-based workflows, including fixing code for a Mastodon client, involving "bug fixes, new features, [and] full builds."
Meta also stated that it used the model to run the high-end video editing software DaVinci Resolve, with results showing that the model can review and trim footage.
Open-source weight models offer several industry advantages over closed and proprietary models. These models are typically more cost-effective while also providing transparency and flexibility that closed-source models cannot achieve.
Meta is currently offering this new model through several different API platforms.