On November 9th, Meta officially launched Muse Ember, the smallest and most efficient high-performance AI model to date. This model is specifically designed for artificial intelligence agents to perform routine, complex tasks on a daily basis, including tool invocation, small coding, request routing, data extraction, and long document processing.
In terms of core performance and cost, Muse Ember shows strong market competitiveness. According to the Artificial Analysis Intelligence Index results, it surpasses Claude Haiku5.5 and GPT-6Luna at every reasoning level.
In the AII benchmark tests, Muse Ember scored 36 points, which is almost equal to that of the larger Muse Spark1.3 (37 points), but the single-task processing cost was reduced to approximately $0.14, far lower than Muse Spark1.3's $2.39. However, in the actual coding test with FrontierCode2.0-AA, Haiku performed better, scoring 34.9%, while Muse Ember scored 27.2%.
In terms of multimodal and input specifications, this model fully supports processing of text, images, audio, and video. The maximum number of tokens in a single context window can reach up to one million. Currently, Muse Ember has been officially launched through the Meta Model API and Muse Code. The pricing for input and output tokens is set at $0.10 per million tokens and $0.50 per million tokens, respectively.
This progress indicates that the AI industry is shifting from focusing on sheer parameter size to improving inference efficiency. Lightweight and cost-effective agent-specific models may accelerate the large-scale deployment of AI agents in specialized business scenarios.