Meta has added another entry to its growing AI portfolio with the official announcement of Muse Spark Open, which the company describes as the most powerful open-weights model it has developed to date. The model is designed for both software development and automation-focused agentic tasks.

Open-weights models give developers more transparency, flexibility and cost advantages than closed systems, and Meta's new release leans fully into that philosophy.

Impressive specs on paper

meta1

Muse Spark Open arrives with striking numbers: 816 billion parameters and a 1 million-token context window. According to Meta, the model outperforms popular open-weights rivals.

Another standout feature is its ability to run locally on Apple computers. But tapping that power isn't easy: local execution requires at least 1 TB of RAM. Meta says the most efficient results were achieved on Mac Studio machines equipped with the M5 Ultra chip.

Tested on real-world tasks

Meta didn't stop at the spec sheet. Muse Spark Open was integrated into a range of agentic workflows via Muse Code, the company's AI coding terminal.

As part of this, the model was put through real-world coding tasks on a Mastodon client, including bug fixing, adding new features and full configuration. It was also reported to have successfully inspected images and performed cutting operations while running DaVinci Resolve.