Meta releases Muse Ember, a lightweight multimodal model built for agent execution
The fourth model in Meta's Muse family is priced at $0.10 per 1M input tokens and is positioned to handle the high-volume work that AI agents repeat all day.
Meta's Superintelligence Lab has released Muse Ember, the fourth and lightest model in its Muse family following Spark, Glimmer and Wonder. The model is explicitly positioned for agent execution work: tool calls, routing, extraction, subtask coding and parsing long documents. It is multimodal across text, images and audio, so an agent can read screenshots, charts or scanned documents without a separate model.
According to composite benchmarks from Artificial Analysis cited by TechCrunch, Ember is cheaper and scores higher than Anthropic's Haiku and OpenAI's Luna, with input pricing of $0.10 per million tokens that puts it in direct competition with open-weight flash-class models such as DeepSeek V4.1 Flash and GLM 5.3 Flash. Meta's framing — the larger Wonder model plans while Ember handles the volume — formalizes a two-tier architecture that is quickly becoming standard in agent stacks. Ember is API-only at launch, with no indication of open weights, which keeps on-premise deployments relevant for data-sensitive workloads.