Muse Ember

Meta launched Muse Ember as its smallest and most efficient model, available in Muse Code and through the Meta Model API as muse-ember-1.0 with Responses API support, at $0.10 per million input tokens, $0.004 per million cached input tokens, and $0.50 per million output tokens. Alexandr Wang said it delivers more intelligence per dollar than Haiku, Luna, and other flash models at any reasoning setting. The model has a 1 million-token context window and supports up to 512,000 output tokens, letting agents read files, run tests, make targeted code fixes, and call tools in sequence across apps while processing text, images, audio, and video natively. At maximum reasoning Ember scores 35.8 on the Artificial Analysis Intelligence Index at about $0.136 per task, versus 36.2 at $0.411 for Haiku, and leads rivals on AutomationBench-AA at 37.8 and Agents' Last Exam at 61.6.

Meta Superintelligence Labs released Muse Ember on November 9, 2026, describing it as the lab's lightest and most cost-efficient model, built for the work agents repeat all day: tool calls, coding subtasks, routing, extraction, and long document parsing. Meta positioned Ember as cheaper and smarter than Anthropic's Haiku and OpenAI's Luna on Artificial Analysis composite scores. Ember is priced at 10 cents per million input tokens, making it cost-competitive with open-source options such as DeepSeek V4.1 Flash and GLM 5.3 Flash. Meta designed it to pair with the larger Muse Wonder in agentic harnesses, saying "Muse Wonder plans, Muse Ember handles the volume." It handles text, images, audio, and video, and is the fourth model under the Muse brand after Spark, Glimmer, and Wonder.

Meta's Muse Ember (meta/muse-ember) is the smallest and most efficient model from Meta, built for complex reasoning and the repetitive work agents do all day, including tool calls, coding subtasks, routing, extraction, and long document parsing. The listing shows a release date of Nov 9, 2026 and a 1M-token context window, with text, image, video, audio, and PDF input returning text output. Measured serving figures on the page show roughly 163 tokens per second throughput and 2.11s P50 latency, with tail latency rising to 8.44s at P99 and 25.59s end-to-end at P99. Reliability figures include a 5.83% tool call error rate and 1.56% structured output error rate, alongside a 76% cache hit rate and 100% uptime, all relevant to agentic integration planning.

Google Cloud's Gemini Enterprise Agent Platform documentation lists Muse Ember 1.0 from Meta with model ID meta/muse-ember-1.0 in Preview stage, released November 9, 2026. It is described as an efficient reasoning model for high-volume agent tasks including tool calls, coding subtasks, routing, extraction, and long-document processing, and can be paired with Muse Wonder 1.0 or Muse Spark 1.3 for planning. The documentation specifies a maximum input of 1,048,576 tokens and maximum output of 524,288 tokens, with text and image inputs producing text output. Function calling, structured output, and reasoning are supported in preview, while batch predictions are not, and requests go to the global publishers endpoint using the bare model ID.

Vals AI lists Muse Ember as a Meta model released November 9, 2026, with a 1M context window and 524k maximum output tokens, supporting text, image, and file inputs. It ranks 13th of 36 models on the Vals Index with a 60.48 percent accuracy score, and its strongest result is second of 37 on HLAB-Vals-1.1. Independent benchmark rows show 60.91 percent on the Vals Finance Index, 75.69 percent on the Vals Legal Index, 72.17 percent on EMB v1.1, 63.83 percent on Finance Agent v2.1, 83.70 percent on Vibe Code Bench v1.1, and 39.39 percent on Terminal-Bench 4.0, with measured cost per test of $0.153 and reported latency around two minutes.