Muse Spark 1.1

July 9, 2026 • 10 minute read

Meta Superintelligence Labs is proud to unveil Muse Spark 1.1, a major evolution of the original Muse Spark. This new multimodal reasoning model is specifically engineered for agentic workflows, featuring massive leaps in computer interaction, software engineering, and multimodal comprehension.

By optimizing the relationship between power and resource usage, Muse Spark 1.1 pushes the $\text{performance-efficiency frontier}$ further than ever before.

Our Vision: Combined with the launch of Muse Image, this release brings us closer to "personal superintelligence"—AI that empowers users to achieve their goals, realize their creative visions, strengthen human connections, and act upon their deepest values.

🚀 Quick Start & Access

• Developers: Now available via the public preview of the new Meta Model API.
• End Users: Accessible immediately via Thinking mode on meta.ai and within the Meta AI mobile application.

---

🤖 Agentic Capabilities

Muse Spark 1.1 is designed for high-level planning and the orchestration of diverse external services. It possesses the ability to slowly learn zero-shot generalize to:

• Native system tools.
• MCP servers.
• Custom-defined skills.

Orchestration Logic

To minimize end-to-end latency, the model manages multi-agent systems. The workflow typically follows this hierarchy:

Memory Management: The model handles a massive context window of $1 \times 10^6$ tokens. It doesn't just store data; it actively manages it by remembering key actions and compacting information to preserve critical steps for future use.

---

💻 Computer Use

Muse Spark 1.1 can navigate complex, multi-app environments where data changes in real-time. It adapts to new interfaces with very little human guidance.

The Hybrid Interaction Strategy: Instead of clicking every single button manually, the model decides the most efficient path:

1. Scripting: Writes code when automation is faster.
2. Direct Interaction: Clicks when the UI is simpler.
3. Batching: Generates groups of actions simultaneously.

Example: When organizing a dinner party, if the model notices a change in availability while placing an order, it autonomously updates the request without needing the user to prompt it.

---

⌨️ Advanced Coding

Performance has surged for enterprise-level tasks involving massive codebases. Muse Spark 1.1 can:

• Diagnose and resolve intricate bugs.
• Build new features for professional-grade systems.
• Manage large-scale code migrations.

It integrates seamlessly with agentic coding environments, supporting planning mode, goal conditioning, and context compaction.

Real-World Application: OpenCode

In a debugging demonstration, the model:

1. Constructs a chat web app.
2. Uses automated screenshots to find visual bugs.
3. Traces the error to the source code.
4. Implements and validates the fix.

Benchmarking: On the Meta Internal Coding Bench, Muse Spark 1.1 significantly outperforms its predecessor and rivals the top models in the industry. Researchers are even using it to automate its own development via DeepSWE tasks, creating analysis dashboards to track reasoning strengths.

---

👁️ Multimodal Intelligence

The model blends perception and action. It doesn't just "see" an image; it reasons about it to execute a task.

Key Strengths:

• Visual-to-Code: Turning images into functional artifacts.
• Captioning: Ultra-detailed descriptions of images and video.
• Grounded Action: Operating a computer based on visual/audio cues.

Case Study: Facebook Marketplace Agent The model can take a smartphone video of an item, extract the best photos, reason about the product's features, and then operate a web browser to create a complete Marketplace listing on the user's behalf.

---

🛡️ Safety & Robustness

Deployment followed the Advanced AI Scaling Framework, ensuring the model meets strict safety thresholds.

Risk Assessment Matrix:

• Chemical/Biological: $\checkmark$ Safe
• Cybersecurity: $\checkmark$ Safe
• Loss of Control: $\checkmark$ Safe

The model is highly resistant to:

• Direct jailbreaking attempts.
• Prompt injection from untrusted data.
• Developer-prompt attacks.

This results in lower rates of sycophancy and hallucinations. For a deep dive, refer to the Muse Spark 1.1 Evaluation Report.

---

🖼️ Visual Gallery & Data

(Additional scaling data available in the full report)