article --summary
The summary
Meta has introduced Muse Ember, a compact and cheaper member of its Muse model family, aimed at high-volume agent subtasks such as reading files, running tests, and making targeted fixes while a larger model handles planning. The piece explains how it is offered through Meta's own tools and API, with weights kept private as with earlier Muse releases.
On pricing and benchmarks, Ember is presented as matching Claude Haiku's score on the Artificial Analysis index at roughly a third of the cost, and undercutting several Chinese rival models on listed rates. The article also notes where Ember lags, including a repository-merge benchmark where Haiku and DeepSeek score higher.
Finally, it covers computer-use results on OSWorld, where Ember roughly matches Haiku on partial credit but completes more tasks outright, and notes that Meta trained the model using its larger Muse Wonder Full model.
Editorial note: This is a summary of an article first published by The New Stack. The full piece is on their site.