Meta Muse Spark 1.1 Release Notes

Meta Superintelligence Labs has introduced Muse Spark 1.1, a multimodal reasoning model specifically engineered for agentic tasks. This release marks a significant upgrade over the original Muse Spark, offering substantial improvements in tool use, computer interaction, coding, and multimodal understanding to advance the performance-efficiency frontier for personal superintelligence.

Agentic Orchestration and Planning

Muse Spark 1.1 is designed to function as a central orchestrator for complex projects, utilizing a multi-agent system to optimize end-to-end latency. The model can gather context, formulate plans, and delegate execution to parallel subagents. When acting as a subagent, it adheres to specific tasks and knows when to escalate issues back to the main agent.

Key agentic capabilities include:

• Zero-shot generalization: The model generalizes to new native tools, MCP servers, and custom skills without prior training.
• Context Management: It manages a 1-million-token context window, allowing it to retrieve information from early work and compact critical steps for later use.

Computer Use and Automation

Muse Spark 1.1 excels in workflows that span multiple applications where information changes dynamically. It navigates unfamiliar interfaces with minimal human intervention and maintains context across extended sessions.

To optimize efficiency, the model employs a hybrid approach to interaction:

• Scripting: The model writes scripts when automation is faster.
• Direct Interaction: It uses clicks and direct interface interaction when simpler.
• Batching: It generates batches of actions at each step to reduce latency.

Advanced Coding Capabilities

Coding performance has been substantially improved for real-world tasks involving large, complex codebases. Muse Spark 1.1 is capable of diagnosing and fixing complex bugs, implementing new features in enterprise-grade systems, and executing large-scale code migrations.

Technical highlights for coding include:

• Integration with Agentic Setups: Support for planning mode, goal conditioning, subagent delegation, and context compaction.
• Multimodal Debugging: In the OpenCode demo, the model combines coding, multimodal understanding, and tool calling to identify user-visible failures via screenshots and trace them back to the relevant code for fixing.
• Internal Benchmarking: On Meta's Internal Coding Bench, Muse Spark 1.1 is competitive with leading frontier models.

Multimodal Reasoning and Perception

Muse Spark 1.1 integrates perception and action, allowing it to operate computers on a user's behalf based on visual and audio inputs. It is proficient in visual-to-code artifact generation, ultra-descriptive image and video captioning, and agentic workflow execution.

For example, in a Facebook Marketplace agent use case, the model can extract photos from a smartphone video and reason about the product to create a listing in a web browser.

Safety and Adversarial Robustness

Following the Advanced AI Scaling Framework, Meta conducted extensive safety evaluations across frontier risk categories, including Cybersecurity, Chemical & Biological risks, and Loss of Control.

Evaluations indicate that Muse Spark 1.1 operates within safe margins and demonstrates strong resistance to:

• Direct jailbreaks
• Indirect attacks from untrusted data
• Prompt injection
• Developer-prompt attacks

This results in lower hallucination rates and reduced sycophancy compared to previous versions.

Availability and Developer Access

Muse Spark 1.1 is now available in "Thinking" mode within the Meta AI app and on meta.ai. For developers, Meta has launched a public preview of the new Meta Model API, which is OpenAI-compatible.

Industry partners have highlighted the model's utility as an agentic foundation:

• Replit: CEO Amjad Masad noted the model's combination of massive context, multimodal support, and parallel tool calling.
• Cline: CEO Saoud Rizwan emphasized the model's strong tool use and price point for running coding workloads at scale.
• Box: VP of AI Products Yashodha Bhavnani stated that the model's enterprise capabilities are competitive with leading frontier models for structured, procedural workflows.