Meta releases Muse Spark 1.1, a multimodal agent with a 1-million-token context window

On July 9, 2026, Meta Superintelligence Labs unveiled Muse Spark 1.1, a multimodal model built for complex agentic tasks with a 1-million-token context window.

What Muse Spark 1.1 can do

The model was designed for multistep workflows in which an agent coordinates several applications and external services at once. Muse Spark 1.1 can act as the main orchestrator of a multi-agent system or as a specialized subagent within one.

• Context window: 1 million tokens
• Launch: July 9, 2026, as a public preview through the Meta Model API
• Available in "Thinking" mode in the Meta AI app and on meta.ai
• Supports images, video and audio
• The API is compatible with the OpenAI format

Its multimodal capabilities include generating code from visual input, writing detailed descriptions of images and video, and running complex workflows that use several tools at the same time.

How does the model handle code and computer control?

On real-world tasks with large enterprise codebases, Muse Spark 1.1 shows a significant improvement over its predecessor: it diagnoses bugs, implements features in existing systems and performs large-scale code migrations. Results were especially strong on web application development and multi-turn question-answering scenarios.

In computer-control scenarios, the model does not limit itself to step-by-step click automation. It decides for itself when to write a script and when to interact with the interface directly, and it generates batches of actions simultaneously. Amjad Masad, CEO of Replit, commented on the launch:

"The most impressive thing about

Muse Spark is how much is packed into one model: a million-token context, full multimodal input (images, video, PDF), built-in search with citations, strong reasoning, top-tier coding capabilities, especially in frontend and design, structured output and parallel tool calls."

Where to run the model

Developers get access to Muse Spark 1.1 through the public preview of the Meta Model API, which has an OpenAI-compatible interface, making migration from other providers easier. End users can already use the model in "Thinking" mode in the Meta AI app and on meta.ai.

Before release, Meta ran a safety evaluation under its Advanced AI Scaling Framework, covering risks in chemical and biological weapons, cybersecurity and loss of control over AI. According to the results, the model was found to operate within safe parameters and to be resistant to jailbreaks and prompt injection attacks.

What this means

Meta is entering the market for advanced agentic AI with a competitive offering: a million-token context, an open API and full multimodality. It is a direct challenge to OpenAI and Anthropic in the enterprise agent segment.

Frequently asked questions

How does Muse Spark 1.1 differ from other agentic models?

According to Meta, the model combines a 1-million-token context window, native multimodality (images, video, PDF, audio) and parallel tool calling in a single system. Cline CEO Saoud Rizwan particularly highlights the attractive balance of capability and cost for real-world agentic workloads.

Do I need a separate API key from Meta?

Yes. Muse Spark 1.1 is available through the Meta Model API, which is an independent provider not affiliated with OpenAI or Anthropic. The interface is compatible with the OpenAI format, which makes it easier to move existing code to the new provider.

*Meta has been designated an extremist organization and is banned in Russia.