Meta Debuts Muse Ember to Power High-Volume Agentic Workflows

Meta Superintelligence Lab has released Muse Ember, the smallest and most cost-efficient model in the Muse lineup, according to reporting from TechCrunch and a Meta blog post. Designed for the repetitive tasks agents perform daily—including tool calls, routing, extraction, coding subtasks, and long document parsing—Ember is positioned as a high-volume execution engine.

Meta describes the model's role within an agentic harness as the worker to a larger model's architect: while the frontier Muse Wonder model establishes rules, Muse Ember applies them across thousands of cases. To support this, Ember features a 1M-token context window and up to 512K output tokens, allowing it to process entire document sets or codebases in one call, Meta reports.

On the Artificial Analysis leaderboard, Meta claims Ember provides more intelligence per dollar than GPT-6 Luna and Claude Haiku 5.5, matching Haiku's top score at roughly one-third of the cost. The model is available via the Meta Model API at $0.10 per million input tokens and $0.50 for every million output tokens. TechCrunch notes that Ember is also cost-competitive with open-source models such as GLM 5.3 Flash and DeepSeek V4.1 Flash.

Early partner testing highlights the model's speed and utility. Dane Knecht, CTO of Cloudflare, stated Ember ran up to 30% faster than comparable models in their tests. Chak Pothina, Inference Lead at You.com, reported that coding quests saw a p95 latency of 2.2 seconds, outpacing leading open-weight models by 4x to 13x. Other partners including Oracle, Box, and Cline also noted Ember's effectiveness in engineering tasks and enterprise content analysis.

Beyond text, Meta trained Ember for multimodal inputs, enabling agents to process voice memos, video, charts, and screenshots without requiring a separate model.