Meta unveils Muse Ember, a low‑cost AI model for agent workloads

Meta’s Superintelligence Lab announced Muse Ember on Monday, positioning it as the company’s most economical model for running AI agents at scale. Priced at ten cents per million input tokens, Ember promises to undercut leading commercial offerings while retaining the ability to perform complex reasoning, tool calls, coding subtasks, routing, extraction and long‑document parsing.

What is Muse Ember and why does it matter?

Muse Ember is the fourth model released under Meta’s Muse brand. According to the TechCrunch release, it is the lightest and most cost‑efficient model in the line‑up, built specifically for the repetitive tasks that agents perform throughout the day. By lowering the per‑token price to $0.10, Ember makes it financially viable to deploy agents that must handle thousands or even millions of interactions, a barrier that has limited many enterprises from adopting agentic workflows at scale.

How does Ember differ from competing models?

The announcement cites Artificial Analysis’s composite benchmark, showing Ember as both cheaper and smarter than Anthropic’s Haiku and OpenAI’s Luna. Its cost‑per‑token also sits competitively alongside open‑source flash models such as DeepSeek V4.1 Flash and GLM 5.3 Flash. Unlike many high‑performing models that focus on a single modality, Ember is trained to accept text, images and audio, allowing it to read screenshots, charts, scanned documents, voice memos, meeting recordings and video without requiring a separate specialist model.

Which use cases is Ember designed for?

Meta frames Ember as the workhorse for volume‑heavy scenarios. A typical pattern involves a larger model—currently Muse Wonder—establishing high‑level policies or reasoning, then delegating the execution of those policies to Ember. In a customer‑support setting, for example, Wonder might decide how to categorize a complaint, while Ember applies that categorization across tens of thousands of tickets. The model’s multimodal capability further expands its reach to tasks such as extracting data from screenshots, transcribing audio memos, or parsing PDFs, all within a single inference call.

How does Ember fit into the broader Muse ecosystem?

Ember follows a staggered release cadence: Spark arrived in April as a mid‑weight model, its open‑source twin Glimmer launched in August, and the frontier model Wonder debuted in October. The consumer‑facing Muse Agent, released in September, already integrates features like Spaces and video chat, and Ember is expected to power the backend of those high‑frequency interactions. Meta describes the relationship as “Muse Wonder plans, Muse Ember handles the volume,” indicating a hierarchical orchestration where the heavyweight model focuses on strategy and Ember on execution.

Practical implications for developers and researchers

For developers building agentic systems, Ember’s low price point reduces the operational budget for large‑scale deployments. Its ability to process multiple modalities means that pipelines no longer need to stitch together separate vision, speech and language models, simplifying architecture and lowering latency. Researchers interested in benchmarking cost‑efficiency can now compare Ember against both commercial and open‑source flash models using the same token‑based pricing metric.

Trying Ember with Grafy

Grafy users can experiment with Ember through the Grafy Chat interface, which lets you run prompts against a variety of models side‑by‑side. By selecting Ember as the active model, you can compare its responses, token usage and multimodal handling against larger models like Muse Wonder or other available providers. For document‑heavy workflows—such as long‑form extraction or PDF parsing—Grafy Reader offers a convenient way to upload PDFs and query them with Ember, letting you see first‑hand how the model performs on long‑document reasoning at a fraction of the typical cost.

---

Source: https://techcrunch.com/2026/11/09/meta-releases-muse-ember-a-new-lightweight-model-for-agent-execution/

---

Source: TechCrunch — AI — Meta releases Muse Ember, a new lightweight model for agent execution