Meta: Muse Spark 1.1 brings multimodal reasoning with one million token context and sub-agent coordination via Model API

On July 9, 2026, Meta Superintelligence Labs released Muse Spark 1.1, a multimodal reasoning model focused on agentic tasks. A reasoning model is a variant of a large language model that conducts an explicit chain of thought before answering, which improves accuracy on complex, multi-step problems.

One million tokens and sub-agent coordination

Muse Spark 1.1 operates with a context window of one million tokens — enough for entire code repositories or a multi-hour history of an agentic task, placing Meta on par with Google Gemini’s offering and surpassing the standard 200,000 tokens of most competitors. The model can coordinate sub-agents and navigate flows across multiple apps, choosing between programmatic automation and direct UI interaction. Meta reports measurable improvements on real coding tasks with large codebases compared to Muse Spark 1.0.

How does Spark fit into the Muse family?

Last week’s Muse Image and Muse Video (released July 7) covered generative modalities, and Spark 1.1 adds agentic reasoning — completing Meta’s entire lineup in one week. Unlike the Llama series, Muse models are not open-weights: they are available through the Meta Model API (public preview) and in the “Thinking” mode of the Meta AI app at meta.ai. This is a strategic shift for a company that spent years building its reputation on open models.

Safety evaluation

Meta reports that Spark 1.1 was evaluated under its internal Advanced AI Scaling Framework — tests cover chemical and biological risks, cybersecurity, and loss-of-control risks, with all results within defined safe limits. The model accepts images, video, and PDF documents as input, which in combination with the million-token context positions it for analyzing extensive business documentation — a segment where it directly competes with Claude and the new GPT-5.6 Sol.