最新模型
Meta Launches Muse Ember, Low-Cost Multimodal Reasoning for High-Volume Agent Subtasks
On November 9, Meta launched Muse Ember, with a million-token context window and adjustable reasoning effort. It targets tool calls, coding subtasks, and document processing. Standard input costs $0.10 per million tokens, while requests with more than 200,000 input tokens are charged at a higher rate.
Meta released Muse Ember on November 9, making it available in Muse Code and the Meta Model API. The model targets repetitive, well-scoped tasks in agent workflows, including field extraction, request classification, tool calls, and localized code changes. Meta recommends having Muse Wonder or Muse Spark handle planning, then delegating high-volume subtasks to Ember. Launch announcement
On the technical side, muse-ember-1.0 accepts text, images, audio, video, and PDFs, and generates text. It has a context window of up to 1,048,576 tokens and a maximum output of 524,288 tokens. The API provides Responses, Chat Completions, and Messages endpoints. Reasoning effort can be adjusted from minimal to max; the default is the highest setting. The model also supports tool calls and structured outputs. Model documentation
Costs depend on prompt length. Meta’s developer announcement lists input/output prices of $0.10/$0.50 per million tokens for requests with up to 200,000 input tokens, and $0.30/$1.50 when that threshold is exceeded. So while the million-token context window can hold more code and documents, long conversations may still trigger higher rates. For multi-agent systems, teams should measure the total tokens, retries, and verification costs required to complete a task. Developer announcement
The evaluations also show capability limits. Meta reported an Ember score of 40.9 on Terminal-Bench 4.0, above Haiku 5.5’s 39.2. However, the two models used different agent execution frameworks, so the scores cannot be treated as a direct comparison of the models themselves. On FrontierCode 2.0-AA, which scores whether code changes are mergeable, Ember scored 27.2, below Haiku’s 34.9. These figures come from Meta’s launch materials and should be retested with the same tools, budgets, and test sets. Evaluation results and conditions
Vercel added the model to AI Gateway the same day under the name meta/muse-ember. It can be called through the AI SDK and several compatible APIs. Engineering teams should next verify its performance on Traditional Chinese document extraction, tool-parameter accuracy, and the latency and success rate when reasoning effort is reduced. Input modalities and output limits should also be checked separately for each provider. Vercel update