Meta releases Muse Ember, a new lightweight model for agent execution

Meta's Superintelligence Lab has unveiled Muse Ember, an AI model that is both lightweight and budget-friendly. The new model was specifically designed for handling complex reasoning tasks and repetitive agent actions such as tool calls, coding subtasks, routing, data extraction, and parsing lengthy documents. According to Meta, Ember's primary strength lies in its cost-efficiency, with a cost of just $0.10 per million input tokens - significantly lower than its competitors like Anthropic's Haiku and OpenAI's Luna, and even on par with open-source alternatives such as DeepSeek V4.1 Flash and GLM 5.3 Flash.

Meta suggests that Ember pairs best with a larger model like Muse Wonder, forming an agentic system. For instance, in a customer support scenario, a heavier model like Wonder could establish rules for handling specific issues, while Ember would then apply these rules across a large number of cases. The model is also trained to handle various types of media, including text, images, and audio.

This flexibility allows agents to work with diverse input types, like screenshots, charts, scanned documents, voice memos, meeting recordings, and videos, without needing separate models.

Since its release, Muse Ember is the fourth model under Meta's Muse brand. Following its predecessor, the mid-weight Spark (released in April), the open-source twin Glimmer (released in August), and the powerful frontier model Wonder (released in October), Ember continues Meta's trend of releasing increasingly capable AI models.

The Muse Agent, Meta's flagship consumer AI product, was launched in September and has been met with positive reception. Recent updates to the agent, such as Spaces and video chat, have further enhanced its appeal.