Meta Launches Muse Ember, Its Smallest Agent Model

Meta released Muse Ember on November 9, and the pitch is blunt. This is not the smartest model in the room. It is the cheap worker you send to do the jobs that pile up when an agent runs all day.

Agents spend most of their tokens on small steps. They call tools, read files, route requests, pull facts from long documents, and fix small pieces of code. Meta says Ember was built for exactly that load, and priced so teams can run it at volume without sending every step to a frontier model.

The Announcement

Meta calls Ember its smallest and most efficient model. Small here means cost and footprint, not context. The model can take in 1,048,576 tokens in one call and return up to 512K output tokens, per Meta. That is enough room for a full codebase or a large set of documents in a single request.

The model is multimodal by default. It reads text, images, audio, and video, so an agent can work from screenshots, charts, scanned documents, voice memos, and meeting recordings without calling a second model. Meta also says Ember is the first model it trained using Muse Wonder Full, the large version of its frontier model.

The intended split is clear in Meta’s own line: Muse Wonder plans, Muse Ember handles the volume. A larger model sets the plan and the rules. Ember then applies them across many cases, tools, and files. We have seen the same split grow in tools such as Perplexity Teammates, where the value is not one clever answer but work that keeps running.

Important Details

Meta published benchmark claims from Artificial Analysis dated November 9. On the Artificial Analysis Index at max reasoning, Ember scored 36, the same rounded score as Claude Haiku 5.5 and ahead of GPT-6 Luna at 28. On AutomationBench-AA, which tests business workflows from start to finish, Meta reports 37.8 for Ember, 20.2 for Luna, and 13.9 for Haiku.

Read those numbers as launch claims, not as settled proof. The table comes from Meta and Artificial Analysis, and Meta chose which rival models and tests to show. Ember also trailed Haiku on one coding test in the same table, FrontierCode 2.0-AA, at 27.2 to 34.9. The honest read is narrower: Ember looks strong on workflow tests and close on general scores, at a much lower list price.

Meta says it tested Ember before release under its scaling framework for chemical and biological risks, cyber risks, and loss of control risks. It says the model stayed below the high risk thresholds in the areas it assessed, and that it got the same core safety training as larger Muse models. Those are Meta statements about its own tests.

What’s Next

Confirmed steps are simple. Ember is live now, it supports the Responses API, and Meta says more efficient models are already in training. No open weights were announced, and Meta did not publish a parameter count in the launch post.

Open questions are about real workloads. Benchmark tasks are tidy. Production agents hit messy files, odd tools, and long chains where one bad step can spoil the rest. Teams will learn more from their own routing logs in a week than from any launch chart.

The takeaway is about price pressure, not bragging rights. When the worker model costs this little, rivals have to justify why their small models cost more for the same repeated work. That pressure lands on Haiku, on Luna, and on every agent platform that passes model costs straight to its users.

Facts checked November 10, 2026, against the Meta announcement dated November 9, 2026.

Sources

• Meta, Introducing Our Smallest, Most Efficient Model: Muse Ember, November 9, 2026: https://research.meta.ai/blog/muse-ember
• TechCrunch, Meta releases Muse Ember, a new lightweight model for agent execution, November 9, 2026: https://techcrunch.com/2026/11/09/meta-releases-muse-ember-a-new-lightweight-model-for-agent-execution/
• SiliconANGLE, Meta introduces Muse Ember to power cost efficient AI agents, November 9, 2026: https://siliconangle.com/2026/11/09/meta-introduces-muse-ember-to-power-cost-efficient-ai-agents/