Muse Spark 1.1

Muse Spark 1.1, released July 9, 2026, is a multimodal reasoning model from Meta built for agentic tasks. It accepts text, image, video, PDF, and audio input, and supports a context window of 1.0M tokens.

Muse Spark 1.1 plans work and then routes it across the tools and services you expose. It holds either position in an agent hierarchy: a main agent that delegates, or a subagent that takes an assignment and reports back. That flexibility matters when you compose a system from several models and want one of them coordinating the rest.

Muse Spark 1.1 works with new tools, Model Context Protocol (MCP) servers, and custom skills without examples. You describe the interface and Muse Spark 1.1 uses it, so adding a tool doesn't mean writing few-shot demonstrations for every call pattern. Parallel tool calling runs independent calls at once instead of in sequence, which shortens the critical path in a multi-tool step.

Structured output keeps responses machine-readable, so a downstream stage parses a schema instead of scraping prose. Built-in search returns citations with its results, which gives you a source to check when an answer depends on current information.

Mixed input types cut preprocessing. A product spec that arrives as a PDF, a bug report with a screen recording attached, and a voice note from a customer call all go into the same request. None of them need a separate extraction step before Muse Spark 1.1 sees them.

Set the model to meta/muse-spark-1.1 and call it with the AI SDK or any supported interface like Chat Completions, Responses, or Messages. AI Gateway routes across Meta with automatic failover. Muse Spark 1.1 supports a context window of 1.0M tokens and completions up to 1.0M tokens per request, at $1.25 per million input tokens and $4.25 per million output tokens.