Meta released what it claims to be the “most powerful” open-source weighting model in the United States, named Muse Spark Open. This model also has one distinctive feature: it is specifically designed for Mac.

The evaluation results released on Thursday Meta show that this new model uses 816 billion parameters and a context window of 1 million token, claiming to have better performance than several other popular open-source weight models.

Muse Spark Open natively supports multimodal capabilities and is designed for complex encoding tasks and proxy-based workloads. It is also designed to run locally on Apple computers. However, running this model locally still requires substantial hardware; Engadget reports that at least 1TB of memory is required. The model performs best on Mac Studio equipped with M5 Ultra chips.

Meta indicates that if you can provide sufficient memory, Muse Spark Open could become an important "mainstay" in coding and other desktop tasks. The company claims that it has paired this model with its own AI coding terminals Muse Code and deployed it in various proxy-based workflows, including fixing code for a Mastodon client, which involves "bug repairs, new features, [and] complete builds."

Meta also indicates that it uses this model to run high-end video editing software DaVinci Resolve. The results show that the model is capable of reviewing and cropping materials.

Open-source weight models have several industry advantages over closed and proprietary models. Such models are usually more cost-effective and also provide transparency and flexibility that closed-source models cannot achieve.

Meta is currently offering this new model through multiple different API platforms.