Back to News
NewsJobsAi

Meta Launches Muse Spark 1.1 for Agent Tasks and Coding

Meta launches Muse Spark 1.1, enhancing AI capabilities for coding and agent tasks, alongside a new paid API for developers.

Mercation Admin
July 9, 2026 3 min read
Meta Launches Muse Spark 1.1 for Agent Tasks and Coding

Meta has launched Muse Spark 1.1, a multimodal AI model that competes with leading models like Opus 4.8 and GPT-5.5. Alongside this release, the company has also introduced a public preview of the Meta Model API, providing external developers with their first paid access to Meta's AI models.

Enhanced Agent Functions

Muse Spark 1.1 is specifically designed for tasks requiring coordination across various applications and services. It can seamlessly integrate with new tools, MCP servers, and user skills without needing prior training.

As the primary agent, it collects context, devises a plan, and allocates tasks to sub-agents. In its sub-agent role, it performs designated tasks and can return control to the main agent when needed.

With a context window of 1 million tokens, the model effectively remembers actions and retrieves information from previous stages while compressing context to retain essential steps for users.

Operating on Desktops

Muse Spark 1.1 is optimized for desktop environments, managing multi-application scenarios under varying conditions. It maintains contextual awareness throughout extended sessions and adapts to unfamiliar interfaces with minimal human input.

Rather than relying on a step-by-step approach, the model autonomously selects the best strategies. For routine tasks, it can generate scripts, while for simpler tasks, it interacts directly through the user interface.

Improvements in Programming

Developers have noted significant advancements in Muse Spark 1.1’s ability to manage large codebases. The model can effectively diagnose complex bugs, integrate features into existing corporate systems, and facilitate large-scale code migrations.

It supports popular coding frameworks, offering features such as planning mode, delegation to sub-agents, goal conditioning, and context compression. Early collaborators include Replit, Cline, and Box.

Multimodal Functionality

This model processes text, images, and video, enabling it to generate code from visual layouts and provide descriptions for images and videos. An example demonstrated the model capturing a product listing on Facebook Marketplace, extracting photos, generating a description, and publishing it through a browser.

Benchmark Performance

In comparative benchmarks, Muse Spark 1.1 outperformed its competitors in agent tests, scoring 88.1 on MCP Atlas, compared to approximately 80 for both Opus 4.8 and GPT-5.5. However, it lagged in coding benchmarks, achieving a score of 59.0 on Terminal-Bench 2.0, falling behind GPT-5.5, Gemini, and Claude Opus 4.8.

Despite this, Meta claims that Muse Spark 1.1 shows considerable improvement over its predecessor and competes well with leading models in many categories.

Safety Measures

Meta performed a safety assessment using its Advanced AI Scaling Framework, confirming that Muse Spark 1.1 operates within acceptable limits across various risk categories. The model has shown resilience against attempts at jailbreaking and other cyber threats, with reduced instances of hallucinations.

API Access and Pricing Structure

The Meta Model API is priced at $1.25 per million input tokens and $4.25 per million output tokens, which is competitive compared to other models. Developers can access $20 in free credits upon registration, and the API is compatible with existing SDK formats from OpenAI and Anthropic.

Currently, the public preview is available only to U.S. developers, while clients can interact with the model through the Meta AI app and website.

Strategic Direction

This launch represents a strategic shift for Meta, moving towards a proprietary model with closed weights, as opposed to its previous focus on open models. Meta's AI director, Alexander Wong, emphasized that enhancing coding capabilities is essential for creating fully functional AI agents. The company is also developing a more powerful model, codenamed Watermelon, though no release date has been set.

The original Muse Spark was introduced in April and was initially available only to select partners. Additionally, Meta launched the Muse Image model in July and provided an early preview of Muse Video.