Another strong player has joined the lightweight model track. On November 10, Meta launched Muse Ember, claiming it is its “smallest and most efficient” model so far, designed specifically for complex reasoning and routine tasks that AI agents must perform repeatedly. It doesn’t compete on parameter scale, but on cost-effectiveness.
Performance scores caught up to Haiku, with only a small balance left on the bill
On the Artificial Analysis ranking, Muse Ember’s overall performance can rival Claude Haiku 5.5, but its single-task cost is only one-third of the latter’s, and it outperforms GPT-6 Luna in multiple tests. For example, in AutomationBench-AA for end-to-end business workflows, it scored 37.8 points, far above Haiku’s 13.9 and Luna’s 20.2; it also leads in areas such as agent programming and long document processing.
Swallow the entire code library at once
With a context window of 1 million Tokens for Ember, up to 512,000 Tokens can be output, meaning a single call can feed a complete code library or a whole stack of documents. The AI can handle hundreds of programming sub-tasks in a single session—reading files, running tests, and fixing bugs simultaneously.
It’s also a native multi-functional tool: it handles text, images, audio, and video. Screenshots, charts, scans, meeting recordings, and videos are all no problem for it. There’s no need to use additional models. In terms of security, it has been evaluated by the Meta SI framework before release, and the risk is below the high-risk threshold. The price is affordable: 0.10 US dollars per million Tokens for input, and 0.50 US dollars for output. It’s already available on the Meta Model API, Google Cloud, and Oracle Cloud.