Abstract:
Meta released the lightweight AI model Muse Ember on November 9, local time. According to data from the third-party evaluation agency Artificial Analysis, the model scored 36 points in the comprehensive intelligence index,
which is the same as Anthropic's Claude Haiku 5.5 highest inference gear, but the cost of a single evaluation task is about one-third of the latter
.
Meta positions Muse Ember as its smallest and most efficient model
, mainly for high-frequency tasks that need to be processed repeatedly by AI agents, including calling tools, performing programming subtasks, request routing, extracting information, and parsing long documents. The model can receive multiple input forms such as text, pictures, audio, video, and PDF, and output the results in text form.
Muse Ember has a context window of about 1 million Tokens, can output up to about 512,000 Tokens, and can process larger code bases or document collections in one request
. The model is now online through Muse Code and Meta Model API, and the model number is "muse-ember-1.0".In terms of price, Muse Ember charges US$0.10 per million input tokens, US$0.004 per million cached inputs, and US$0.50 per million output tokens. Tests by Artificial Analysis show that
its highest inference gear costs about US$0.14 per task, and Claude Haiku 5.5 costs about US$0.41 to complete similar tasks
.However, there are differences in the results of different evaluation projects. Tests published by Meta show that Muse Ember is ahead of Claude Haiku 5.5 and OpenAI's GPT-6 Luna in some agents, tool calls and multi-modal tasks; in the FrontierCode 2.0 test that examines whether code modifications can be adopted by project maintainers, Muse Ember's score is lower than Claude Haiku 5.5.
Comments