Meta: Muse Spark Open

Meta: Muse Spark Open is a text model from Meta, released on 15 October 2026. It is described as a coding-optimized open-weight model, trained specifically for agentic coding workflows. That focus shows up in the stated goals of higher first-attempt accuracy, reliable tool calling, and deep understanding of codebases across the repositories developers work in. If you are building developer tooling or automating work inside an existing repo, this model is aimed directly at that setting rather than at general chat alone.

On the input side, Muse Spark Open accepts text, images, video, files, and audio, while its output is text only. The context window is 1,048,576 tokens, and the maximum output is 943,718 tokens. That combination of a very large context and a very large output budget means you can feed in long sources, such as large files or extended transcripts, and still leave room for a substantial generated response. The multimodal input support is worth noting for coding work: screenshots, design files, recorded walkthroughs, and audio notes can all be part of the same request as the surrounding text.

The practical uses follow from those capabilities. Agentic coding workflows are the clearest fit, since the model is trained for tool calling and codebase understanding, which suits tasks like navigating a repository, planning multi-step changes, and invoking tools as part of a loop. The large context helps when a task depends on many files or a long history of prior steps. Because images, video, audio, and files are accepted as input, you can also build flows that turn a screen recording, a UI mockup, or a spoken bug report into text output such as a written summary, a patch description, or a structured report. Reasoning controls and structured outputs make it usable in pipelines where the result has to be parsed rather than read by a person.

You call the model through api14's OpenAI-compatible API using the model id "meta/muse-spark-open". The supported request parameters are frequency_penalty, include_reasoning, logit_bias, logprobs, max_tokens, min_p, presence_penalty, reasoning, reasoning_effort, repetition_penalty, response_format, seed, stop, structured_outputs, temperature, tool_choice, tools, top_k, top_logprobs, and top_p. In practice that means you can tune sampling with temperature, top_p, top_k, and min_p, apply penalties, cap the response with max_tokens, constrain output with response_format or structured_outputs, and wire up function calling with tools and tool_choice. The reasoning and reasoning_effort parameters, along with include_reasoning, let you control how much internal reasoning is exposed or requested, and seed plus logprobs or top_logprobs support reproducible or inspected generations.

One limit to keep in mind is that output is text only, so any task that needs generated images, audio, or video will require a different model. The model is also explicitly coding-optimized, so for work that has nothing to do with code or agentic tool use, a general-purpose text model may be a simpler choice.