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

published: 2026-11-09T18:05:25.121Z
topic: infrastructure
source: https://techcrunch.com/2026/11/09/meta-releases-muse-ember-a-new-lightweight-model-for-agent-execution/
url: https://aiagentnews.io/stories/meta-releases-muse-ember-a-lightweight-model-for-agent-execution

Meta Superintelligence Labs has released Muse Ember, which it describes as the lab's lightest and most cost-efficient model so far. Meta positions it for high-volume agent work such as tool calls, coding subtasks, routing, extraction and long document parsing.

Meta Superintelligence Labs released Muse Ember on Monday, calling it the lightest and most cost-efficient model the lab has produced. Meta says the model is built for complex reasoning execution and for repetitive agent work, including tool calls, coding subtasks, routing, extraction and long document parsing. The company compares its cost per task against Anthropic's Haiku and OpenAI's Luna, using Artificial Analysis's composite benchmark score.

Meta puts Ember's price at 10 cents per million input tokens, which it says is well below Haiku and Luna and competitive with open-source models such as DeepSeek V4.1 Flash and GLM 5.3 Flash. Meta suggests pairing Ember with a larger model, Muse Wonder, in which the bigger model writes the rules and the smaller one applies them across many cases. Meta also says Ember handles text, images and audio, so agents can read screenshots, charts and scanned documents, or work from voice memos and video, without a separate model.

Ember is the fourth model in Meta's Muse line, following the mid-weight Spark in April, its open-source counterpart Glimmer in August, and the Wonder frontier model in October.

## Why it matters if you build agents

Developers running high-volume agent workloads now have a low-priced option from Meta to benchmark against Haiku, Luna and open-source models on cost per task.

The suggested pattern of a larger model planning and a smaller one executing across many cases is one to test in customer-support or batch workflows.

Cost and accuracy claims come from Meta and a third-party benchmark, so teams should verify them on their own tasks before switching.