Meta: Muse Spark 1.3

Meta: Muse Spark 1.3 is a multimodal reasoning model from Meta, built for long-running agentic, multi-agent and coding workflows. It is a text model in the sense that its output is text, but it accepts more than text on the way in. Meta positions it around keeping track of information across extended tasks and working through problems over time, which is why the agentic framing matters: this is a model meant to sit inside a loop rather than answer a single isolated question.

On the input side, Muse Spark 1.3 accepts text, images, video and files, and it returns text. The context window is 1,048,576 tokens, with a maximum output of 943,718 tokens. That combination of a very large input budget and a very large output budget is unusual and shapes what the model is good for. It can take in long documents, media and accumulated conversation history, and it can produce long structured answers, plans or code without being cut off early. Reasoning is a first class part of the model: the supported parameters include reasoning, reasoning_effort and include_reasoning, so you can control how much internal work happens and whether that reasoning is surfaced back to you.

The practical use cases follow from those capabilities. Long-horizon agents that need to hold state across many steps are a natural fit, as are multi-agent setups where one model coordinates or reviews the work of others. Coding workflows benefit from the large output budget, since generated code, diffs and explanations can be returned in one response. Because images, video and files are accepted as input, the model can also handle tasks that mix media with text, such as reviewing a screen recording alongside a bug report, or working through a document plus its supporting diagrams. Anything that requires reading a lot of material and then writing a lot back is where this model is aimed.

You call it through api14's OpenAI-compatible API using the model id "meta/muse-spark-1.3". 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 means you can tune sampling with temperature, top_p and top_k, discourage repetition with repetition_penalty, cap the response with max_tokens, and steer reasoning depth with reasoning_effort. For tool use, tools and tool_choice are available, and for machine-readable replies you can use response_format or structured_outputs. If you want the model's reasoning returned in the response, set include_reasoning.

The main thing to weigh is that this is a text-output model. If your task needs generated images, audio or other non-text output, Muse Spark 1.3 will not produce it, even though it can read images, video and files as input. The very large context and output limits also mean requests can be expensive in latency and resource terms, so for short, simple, high-volume calls a smaller and faster model may be the better choice. Reach for Muse Spark 1.3 when the task is long, multimodal on the input side, agentic, or code-heavy, and when you need room to think and to write at length.