Meta: Muse Spark 1.1

Meta: Muse Spark 1.1 is a multimodal reasoning model from Meta, identified on api14 as meta/muse-spark-1.1. Meta positions it for agentic tasks, meaning work where the model is expected to plan, call tools and carry a task through several steps rather than answer a single question. It sits in the text category of the catalogue, and it was released by the provider on 16 July 2026.

The model accepts text, images, video and files as input, and returns text. That combination makes it a reasoning model you can point at more than plain chat: you can hand it a screenshot, a clip or a PDF document alongside your instructions. The context window is 1,048,576 tokens, so long documents or extended conversations fit in a single request. The maximum output is 943,718 tokens, which is unusually large and matters when you want a full report or a long generated artefact in one call rather than a short reply.

Those capabilities suggest a few practical shapes of work. Document and video understanding: send a contract, a research paper or a recording and ask for a summary, extraction or question answering over the whole thing. Agentic pipelines: define tools, let the model choose when to call them, and let it reason across the intermediate results. Structured extraction: pull fields out of mixed inputs into JSON for downstream systems. Long-horizon generation: produce lengthy text, such as a draft specification, from a large body of source material. Because the output side is text only, use it where the deliverable is written or structured text, not where you need generated images or audio.

You call it through api14's OpenAI-compatible API using the model id meta/muse-spark-1.1, so existing OpenAI SDKs and clients work with a changed base URL and model name. The request parameters it supports are include_reasoning, max_tokens, reasoning, reasoning_effort, repetition_penalty, response_format, structured_outputs, temperature, tool_choice, tools, top_k and top_p. Reasoning behaviour is controlled through the reasoning, reasoning_effort and include_reasoning parameters, tool use through tools and tool_choice, and output shape through response_format and structured_outputs. Sampling is tuned with temperature, top_p, top_k and repetition_penalty, and length with max_tokens.

Two things to keep in mind when deciding. The output modality is text only, so if your task needs generated images or speech, this model is not the fit. And the context and output ceilings are generous but not infinite, so extremely large inputs still need trimming or chunking. For straightforward, short text tasks, a smaller and faster text model may be simpler to operate; Muse Spark 1.1 makes most sense when you need multimodal input, tool use or very long outputs in one request.