Meta introduces muse ember, a new efficient model for AI agent operations

Meta Superintelligence Lab has unveiled Muse Ember, its most lightweight and cost-effective AI model to date. This new model is engineered for the routine, complex reasoning tasks performed by AI agents, such as tool invocation, code subtasks, data extraction, and parsing lengthy documents. Ember is positioned as a highly economical option, priced at 10 cents per million input tokens. Artificial Analysis's composite benchmark places Ember as both cheaper and more capable than rivals like Anthropic's Haiku and OpenAI's Luna, while also being cost-competitive with open-source alternatives such as DeepSeek V4.1 Flash and GLM 5.3 Flash. Designed to complement larger systems, Ember is intended to work alongside models like Meta's Muse Wonder in agentic workflows; for example, a heavier model could set rules while Ember handles high-volume processing. Furthermore, the model supports multimodal inputs, allowing agents to process text, images, and audio—such as voice memos or scanned documents—without requiring separate models. Ember is the fourth release under the Muse brand, following Spark, Glimmer, and Wonder.