Meta: Muse Wonder

Meta: Muse Wonder is a frontier reasoning model from Meta, built for complex personal and professional agentic tasks. It sits in the text category, and Meta positions it around reasoning-heavy work rather than simple chat. If you are choosing a model for an agent that has to plan, reason through intermediate steps, and act on what it finds, this is the class of model Meta intends Muse Wonder to serve.

The model accepts text, images, video, audio, and PDF documents as input, and returns text. That combination matters for agentic work, because a single request can carry a screenshot, a recording, a document, and written instructions together, and the model reasons over all of it. The context window is 1,048,576 tokens, with a maximum output of 943,718 tokens. The large output ceiling is unusual and worth noting: tasks that produce long transcripts, extended analyses, or lengthy generated documents can complete in one response rather than being stitched together across calls.

Those capabilities point to a fairly specific set of uses. Document-heavy analysis is a natural fit, since PDFs can go in directly alongside questions about them. Support and operations agents can take screenshots or screen recordings as part of a ticket and reason about what went wrong. Media review workflows can pass video or audio in and get written output back, which suits summarisation, annotation, or compliance review. Because tool calling is supported, Muse Wonder also works as the reasoning core of an agent that decides which tools to invoke and in what order, including multi-step tasks that mix retrieval, file handling, and generation.

You call it through api14's OpenAI-compatible API using the model id "meta/muse-wonder". Because the interface follows the OpenAI schema, existing clients and SDKs generally work with a change of base URL and model id. The supported request parameters are include_reasoning, max_tokens, reasoning, reasoning_effort, repetition_penalty, response_format, structured_outputs, temperature, tool_choice, tools, top_k, and top_p. In practice that covers sampling control (temperature, top_p, top_k, repetition_penalty), output length (max_tokens), structured responses (response_format, structured_outputs), tool use (tools, tool_choice), and control over how much reasoning the model exposes or performs (reasoning, reasoning_effort, include_reasoning).

One limit to keep in mind is that output is text only. If your application needs generated images, audio, or video, Muse Wonder is not the right component for that part of the pipeline, though it can still reason over those formats as inputs. It is also a reasoning model, so for trivial classification or short deterministic transformations, a smaller and faster text model may be a better fit and cheaper to run at volume. Use Muse Wonder where the task genuinely benefits from long context, mixed media input, or multi-step reasoning with tools.