VMTech
Discuss a project →

Meta Introduces Muse Ember for High-Volume Agent Tasks

Meta Introduces Muse Ember for High-Volume Agent Tasks

Meta Superintelligence Lab has released Muse Ember, describing it as the lightest and most cost-efficient model in its Muse family. Priced at 10 cents per million input tokens, Ember is intended for complex reasoning execution and recurring agent workloads, including tool calls, coding subtasks, routing, extraction and long-document parsing.

Meta used Artificial Analysis’s composite benchmark score to position Ember as both cheaper and smarter than Anthropic Haiku and OpenAI Luna. The company also said its pricing is competitive with open-source models DeepSeek V4.1 Flash and GLM 5.3 Flash.

A lightweight model for repeated execution

Ember is designed for deployments where an agent must apply established instructions across a large number of cases. Meta illustrated this with customer support: a larger model can establish rules for handling an issue, while a lighter model applies those rules across tens of thousands of requests.

Meta summarised that division of work as “Muse Wonder plans, Muse Ember handles the volume.” The design places Ember alongside the company’s higher-capability model rather than presenting it as a replacement for frontier-scale reasoning.

Multimodal inputs without a separate model

Meta said Ember was trained to work across text, images and audio. In the company’s description, agents can interpret screenshots, charts and scanned documents, as well as work from voice memos, meeting recordings and video, without requiring a separate model for those inputs.

The release is the fourth under the Muse brand. It follows the mid-weight Spark model, its open-source twin Glimmer and the frontier Muse Wonder frontier model for planning-oriented work while Ember is positioned to execute repetitive tasks at volume.

What the release means for deployments

For organisations building agent workflows, the announcement highlights a split between planning and execution. Teams can assess whether repeatable support, extraction, routing and document-processing work can be assigned to a lower-cost model, while retaining a larger model for rule-setting or more demanding cases.

That approach makes token pricing, multimodal input needs and the volume of repeatable work practical criteria when selecting models for an agent architecture.

#metaai#aiagents#aimodels#automation
Open analytics
On the site 2 views
min read 3 09.11.2026
Instagram

Meta Introduces Muse Ember for High-Volume Agent Tasks

Open the post on Instagram ↗