Meta launches Muse Spark 1.1 and unveils its new API: committing (even more) to Agentic AI

If all the launches Meta made this week felt like too few, rest assured: apparently, there is more. The company has officially launched Muse Spark 1.1, an enhanced version of its multimodal reasoning model specifically designed to carry out agentic tasks.

But that is not all: so that the tech community can fully tap into its potential, Meta has also announced the public testing launch of its new Meta Model API.

What changes in Muse Spark 1.1

The model can plan, coordinate, and execute complex projects by interacting with different applications and external services. As we explained in this article on the launch of the Muse Spark models, each generation in the Muse series will validate the advances of the previous one before moving on to larger models. This is something we can see in its evaluation report:

And there, we can indeed see that Muse Spark 1.1 significantly outperforms its predecessor across the board. It also outpaces other very well-known AI models on the market. Its closest competition comes from Claude Opus 4.8 and GPT-5.5.

The key features that make Meta’s new model stand out:

• Multi-agent system orchestration: the model can act as a lead agent that designs a strategic plan and delegates the execution of specific tasks to sub-agents working in parallel, optimizing time. In the same way, it also knows how to act as a sub-agent and alert the main agent if it needs help.
• It has a massive memory that allows it to remember actions and retrieve information from very early stages of a work session, compressing data to retain what matters most.
• Advanced computer use: the AI can autonomously navigate unfamiliar interfaces during long sessions. What stands out most to us is that the model knows when it is better to write a script to automate a process and when it is more efficient to interact directly by clicking through the visual interface. As shown in this demo video.

Writing code and “seeing” the screen to fix errors

One of the sectors that will benefit most from Muse Spark 1.1 is software development. The model has been trained to integrate into agentic coding tools and can handle large-scale tasks in complex codebases, such as migrations or resolving difficult bugs.

Thanks to its multimodal capabilities, the AI can combine coding with vision. For example, in a demo within the OpenCode platform, the model built a web application, took automatic screenshots to identify errors visible to the user, traced those errors back to the corresponding lines of code to fix them, and validated the changes on its own. According to Meta, the model is already used daily by its own engineers and has proven competitive against the market’s leading alternatives in its internal benchmark (Meta Internal Coding Bench).

Multimodal capabilities and ad creation in Marketplace

Marketers, take note: if you create ads on Meta, you are going to like this. The model stands out in generating code from visual elements, creating ultra-descriptive captions for photos and videos, and executing multimodal workflows.

One real-world example of its application is the Facebook Marketplace assistant: the model can process a video recorded on a smartphone, extract useful photos, reason about the product’s characteristics, and autonomously open the user’s browser to fill out and publish the listing on the platform. What is truly surprising is that while you are recording, you can see how the model analyzes everything happening in the video in real time to gather enough information before publishing the listing.

Initial reactions to Muse Spark 1.1

For those who want to directly test the model’s reasoning capabilities, Meta has already integrated Muse Spark 1.1 under “Thinking” mode within the Meta AI app and on the meta.ai website.

Meanwhile, developers can already access the model through the new Meta Model API. The first technology partners that received priority access to the system have highlighted its competitive advantages:

• Amjad Masad (CEO of Replit): He points out that the model is a complete agentic foundation that brings together, in a clean OpenAI-compatible package, features such as full multimodal support (images, video, PDFs), built-in web search with citations, structured outputs, and parallel tool calling.
• Saoud Rizwan (CEO of Cline): He emphasizes the value that such robust tool use brings to agentic coding at a price point that makes it viable to run real workloads at scale.
• Yashodha Bhavnani (VP of AI Products at Box): She states that, after testing it with her company’s evaluation sets, the model delivers capabilities that are competitive with today’s frontier models, making it highly attractive for procedural workflows in professional sectors.

As for safety, Meta says it has thoroughly evaluated the model under its Advanced AI Scaling Framework, confirming that it operates within safe margins against cyber, biological, and loss-of-control risks, while showing a lower hallucination rate and greater resistance to prompt injection attacks.

Image: Meta