- Meta released Muse Spark Open, an open-weight AI model whose internal weights are public. Meta claims it is the 'most capable' open-weight model in the United States.
- The model size is 816 billion parameters, and the amount of input it can process at once is 1 million tokens. The claim that it outperforms other popular open-weight models is based on evaluation results published by Meta.
- It was reported to be designed to run locally on Apple computers. However, reports citing Engadget say local use requires at least 1 terabytes of RAM, and the Mac Studio with the M5 Ultra chip was mentioned as the best-suited device.
- Meta cited examples of using its AI coding terminal Muse Code, including bug fixes and feature additions to a Mastodon client, and said it reviewed and trimmed video in the DaVinci Resolve video editing program. All of these examples are Meta's own claims.
O: For developers who want to try local coding assistance on a high-end Mac, the Muse Code case in the article is a useful reference when comparing repetitive bug-fixing work against existing tools.
X: Local operation requires a minimum of 1 terabytes of RAM, and the performance advantage is a claim based on evaluation results published by Meta, so independent verification has not been confirmed.