21:24Machine-translated from Azerbaijani by AIRead in Azerbaijani
Meta Unveils 816-Billion-Parameter Muse Spark Open AI Model
On October 15, Meta publicly released the weights that determine how an artificial intelligence model called Muse Spark Open works. The model has a total of 816 billion parameters, but only 42 billion of them are activated for each task. Muse Spark Open processes different types of data, such as text and images. The model has a 1-million-token context window, meaning it can work with a large amount of data at once. Meta calls it the most powerful open-weight model in the US.
Muse Spark Open is intended for programming, developing complex software, and tasks that require multi-step reasoning. The model can carry out successive steps of a task on its own without continuous human intervention. According to Meta’s tests, the model outperformed GLM 5.3 and Kimi K3 on such tasks and in visual reasoning tests. However, these results were reported by the company itself.
The model has been optimized for Apple Silicon and can run on-device on Mac Studio computers with M5 Ultra processors without sending data over the internet. It uses MLX, the open-source computing framework developed by Apple. The graphics processor in each Mac Studio works with 512 GB of unified memory. As a result, 2 computers can hold the entire model in memory. Together, the devices operate as a single computer with 1 TB of memory and a memory bandwidth of 1.2 TB/s.
The model can also be run on 4 M5 Ultra Mac Studio computers, each with 256 GB of memory. These devices cost a total of approximately $38,000 (approximately 64,600 manats). Apple has not yet announced the exact price or release date of the version with 512 GB of memory. The company has only said that this version will go on sale at the end of October.
Meta has also tested the model on various real-world tasks. Muse Code was used to fix bugs in the open-source Ice Cubes app, add new features, and develop a complete iOS version. In another test, the model operated DaVinci Resolve entirely on its own from start to finish. It reviewed the video frame by frame, selected suitable sections, and converted it to a vertical format.
Muse Spark Open is currently available for download from the Hugging Face platform. Because the model runs locally, it can perform programming and image-related tasks without sending data to a cloud service. Meta believes it is important for US companies not to fall behind Chinese laboratories in developing open-weight models. However, because of its 1 TB memory requirement, the model is better suited to large organizations and researchers than to ordinary users.