Key Highlights

• Meta has introduced Muse Spark 1.1, an upgraded multimodal reasoning model built for agentic AI tasks.
• The model features a 1 million-token context window, stronger coding capabilities, and improved computer use.
• Meta has also launched a public preview of the new Meta Model API for developers.
• Muse Spark 1.1 is now available in “Thinking” mode through the Meta AI app and Meta AI website.

Meta has unveiled Muse Spark 1.1, its latest multimodal reasoning model designed to power more capable AI agents across coding, computer use, planning, and complex workflows. Alongside the launch, Meta has also introduced a public preview of the Meta Model API, allowing developers to build applications using the new model.

The release represents a significant upgrade over the original Muse Spark and follows the recent debut of Muse Image. Together, the launches reflect Meta’s broader ambition to develop what it describes as “personal superintelligence” capable of helping users create, plan, solve problems, and complete real-world tasks.

What is Muse Spark 1.1?

Muse Spark 1.1 is Meta’s newest multimodal AI model built specifically for reasoning-intensive and agentic workloads.

Unlike traditional AI assistants that primarily answer questions, the model is designed to plan tasks, coordinate multiple AI agents, use software tools, interact with computers, write code, and process images and videos within a single workflow.

Meta says the model delivers major improvements in coding, multimodal understanding, tool use, and long-running task execution compared to its predecessor.

The company has also optimized Muse Spark 1.1 for better performance while improving efficiency, enabling more powerful AI capabilities without proportionally increasing computing costs.

Why is agentic AI the focus?

One of the biggest upgrades in Muse Spark 1.1 is its focus on agentic AI.

According to Meta, the model can independently organize complex projects by gathering information, creating execution plans, and distributing work across multiple AI subagents.

As the primary agent, it coordinates the overall workflow. Meanwhile, subagents handle specialized tasks before reporting results back to the main system.

Meta says the model can also work with native tools, Model Context Protocol (MCP) servers, and custom developer skills without requiring additional training.

This allows AI systems to automate increasingly sophisticated workflows across multiple applications and services.

How does the 1 million-token context help?

Muse Spark 1.1 supports an enormous 1 million-token context window, allowing it to process and remember significantly larger amounts of information during extended sessions.

Instead of forgetting earlier conversations, the model actively manages its context by retrieving important details, remembering previous actions, and compressing less critical information while preserving key reasoning steps.

For developers building long-running AI agents, this capability enables more consistent performance across complex projects that may span hours or even days.

What improvements does Meta claim for coding?

Meta says coding represents one of the largest performance gains in Muse Spark 1.1.

The model can diagnose software bugs, implement new product features, migrate enterprise codebases, and build complete web applications. Rather than simply generating code snippets, it is designed to work within large production environments and multi-turn development workflows.

Meta also says the model adapts to modern AI coding environments by supporting planning modes, goal conditioning, subagent delegation, and context management. Internally, the company reports that researchers and engineers are already using Muse Spark 1.1 for software development and AI research workflows.

The model can even automate evaluation tasks by testing AI systems and generating analytical dashboards from the results.

Can Muse Spark 1.1 operate computers?

Another major addition is improved computer-use capability.

Meta says Muse Spark 1.1 can operate across multiple desktop applications while adapting to changing information during a task.

Instead of interacting with every interface manually, the model determines when it is faster to automate actions through scripts and when direct interaction is more appropriate.

The company demonstrated examples in which the model managed online ordering workflows and automatically adjusted its actions when new information became available.

This approach reduces unnecessary human intervention while making AI assistants more capable of completing real-world digital tasks.

How does multimodal AI improve the experience?

Muse Spark 1.1 combines reasoning with image, video, and audio understanding.

Meta says the model can inspect visual content, preserve details across lengthy workflows, and use that information while operating software on behalf of users.

One demonstration showed the model extracting usable product images from smartphone video before automatically creating a Facebook Marketplace listing.

The company also highlighted strengths in visual-to-code generation, highly descriptive image and video captioning, and multimodal workflow automation.

These capabilities allow the model to move beyond text and interact with digital environments using multiple forms of information simultaneously.

What about safety?

Meta says it conducted extensive safety testing before launching Muse Spark 1.1. According to the company, the model was evaluated across frontier AI risk categories including cybersecurity, chemical and biological misuse, and loss-of-control scenarios.

Meta also claims the model demonstrates stronger resistance to jailbreak attempts, prompt injection attacks, and malicious instructions compared to previous versions. The company says Muse Spark 1.1 also reduces hallucinations and shows lower levels of sycophantic behavior.

Detailed findings have been published in the Muse Spark 1.1 Evaluation Report.

Where is Muse Spark 1.1 available?

Muse Spark 1.1 is now available through the Meta AI app and Meta AI website using the new Thinking mode. Developers can also access the model through the newly launched Meta Model API, currently available as a public preview. Meta says early partners are already using the model for enterprise-scale AI workloads that combine long-context reasoning, coding, and autonomous task execution.

Final thoughts

With Muse Spark 1.1, Meta is doubling down on the future of agentic AI. The new model combines stronger reasoning, multimodal understanding, computer-use capabilities, advanced coding performance, and a massive 1 million-token context window into a single platform. As competition among frontier AI models intensifies, Muse Spark 1.1 positions Meta as another major player focused on building AI systems capable of planning, acting, and completing increasingly complex real-world tasks.