Meta releases its strongest AI agent model, Muse Spark 1.1! It supports a million-token context, and is proficient in coding and autonomous computer control.

According to Meta's official announcement, its Superintelligence Labs officially launched the new multimodal reasoning model Muse Spark 1.1 today (9th). This model, specifically designed for "agentic tasks," features a massive 1-million-token context window and showcases remarkable capabilities in multi-agent collaboration, autonomous computer control, and enterprise-level debugging. Meta declares it will leverage this technology to sprint toward a new era of "personal superintelligence."

(Previous: Meta Plans Mass Production of Self-Developed AI Chip Iris in September! Partnering with TSMC to Expand "14-Gigawatt" Computing Empire) (Background: Meta Launches AI Image Generation Tool Muse Image, but Using Your IG Photos Doesn't Require Your Consent)

Table of Contents

Toggle

• Built for "Agentic Tasks," Orchestrates Multi-Agent Collaboration
• Like a Real Human Assistant: Autonomous Computer Control and Advanced Coding
• Visual Perception Evolution, Moving Toward "Personal Superintelligence"

At a critical moment when generative AI is expanding into physical tasks and automated workflows, Meta once again demonstrates its formidable technical moat to the market.

On July 9, 2026, Taiwan time, Meta announced the launch of the latest multimodal reasoning model developed by Superintelligence Labs — Muse Spark 1.1. This is not only a major upgrade from its predecessor but also marks the official transition of artificial intelligence from single-text interaction into an "Agent" era endowed with high-level action and planning capabilities.

Built for "Agentic Tasks," Orchestrates Multi-Agent Collaboration

The core highlight of Muse Spark 1.1 is its significant advancement of the "performance-efficiency frontier." In complex work scenarios, the model demonstrates strong "zero-shot generalization" ability, seamlessly adapting to new native tools, MCP servers, and custom skills.

More revolutionary is its "multi-agent collaboration" architecture. Muse Spark 1.1 can be trained as a "master agent" that orchestrates the overall situation, responsible for gathering context, formulating plans, and delegating tasks to sub-agents in parallel. At the same time, it can also act as a dedicated "sub-agent" focused on a single responsibility, intelligently determining when to escalate difficult issues. Coupled with its massive 1-million-token context window, the model can easily retrieve historical records, actively memorize, and compress key steps, significantly reducing end-to-end system latency.

Like a Real Human Assistant: Autonomous Computer Control and Advanced Coding

On the operational level, Muse Spark 1.1 endows machines with the flexibility of human thinking. In terms of "Computer Use," it excels at executing long workflows across multiple applications and navigating unfamiliar user interfaces with minimal human intervention. For example, when tasked with organizing a dinner party, if an unexpected change occurs during the reservation process, the model can proactively detect it and update decisions in real time, completely without human intervention.

In coding, which developers care about most, Muse Spark 1.1 has demonstrated strong capabilities in diagnosing complex bugs, implementing new enterprise-level features, and executing large-scale code migrations. The model is already being widely used by engineers and researchers within Meta, significantly accelerating development processes. Early partners such as Replit, Cline, and Box have given high praise to its coding abilities and agent infrastructure.

Visual Perception Evolution, Moving Toward "Personal Superintelligence"

As a top-tier multimodal model, Muse Spark 1.1 also excels in vision-to-code generation and extremely detailed image and video descriptions. In an official demonstration on Facebook Marketplace, the agent only needed a video captured by a phone to automatically extract product images, analyze product details, and autonomously operate the browser to complete the listing for the user.

Meta emphasized that Muse Spark 1.1 will strongly complement Muse Image (and Muse Video later this week), signaling the company's full-speed sprint toward the vision of "Personal superintelligence." Currently, developers can access the model through the public preview of the Meta Model API, and general users can also experience it in the "Thinking" mode of the Meta AI App and on the meta.ai website. To ensure safety, the official statement also indicated that the model has undergone rigorous evaluations following the advanced AI extension framework, significantly reducing hallucination rates and possessing strong resistance to adversarial attacks.