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

Meta introduced Muse Ember on November 9, 2026, describing it as the company's smallest and most efficient model for repeated agent and coding work. The model became available the same day through Muse Code and the Meta Model API at $0.10 per million input tokens, $0.50 per million output tokens, and $0.004 per million cached input tokens, with a 1M-token context window and up to 512K output tokens.

Muse Ember is built for the repetitive layer of agent activity: tool calls, coding subtasks, routing, extraction, and long-document parsing. Meta says agents spend most of their tokens on these small, structured steps, and Ember can retrieve records, call tools in sequence, and update systems across applications to execute a business workflow from start to finish. The model natively processes text, images, audio, and video, allowing agents to work from screenshots, charts, scanned documents, voice memos, meeting recordings, or video without a separate model.

Meta positions Ember as the first model trained using Muse Wonder Full, rounding out a portfolio that includes Muse Wonder for the hardest problems, Muse Spark for coding and agentic workflows, and Muse Ember for repeatable versions of those tasks at scale. The company says Ember matches Claude Haiku 5.5's best Artificial Analysis Intelligence Index score at roughly one-third of the cost per task and beats GPT-6 Luna at every reasoning setting, with no model on the full Artificial Analysis leaderboard scoring higher for less per task. Ember scores 36 on the Intelligence Index versus 37 for Muse Spark 1.3, while costing $0.14 per Artificial Analysis task compared to Spark's $2.39.

Early testing partners reported performance gains. Cloudflare CTO Dane Knecht said Ember ran up to 30% faster at the tail than the comparable model his company tested and handled investigations spanning multiple repositories well. You.com inference lead Chak Pothina reported 95th-percentile latency of 2.2 seconds, with Ember completing every coding quest at the lowest latency among models tested and producing 36–48% fewer output tokens.

Meta evaluated Ember under its Superintelligence Scaling Framework for chemical and biological, cybersecurity, and loss-of-control risks, reporting the model stayed below high-risk thresholds across categories assessed. The model received the same core safety training and testing as larger models. Agent-behavior evaluations covered respect for user intent, sensitive-information handling, resistance to hidden instructions in web pages or tool outputs, and seeking human input before high-impact actions.

The release follows Meta's introduction of Muse Wonder a few weeks earlier, described as its most capable model series yet with state-of-the-art results on personal and professional agentic tasks. Meta says the broader goal is to give more people access to more capable superintelligence by transferring key Muse Wonder capabilities into the services they already use, including Active Memory and Active Focus systems that improve context retention and task persistence.