Meta unveils Muse Spark: an ambitious yet imperfect AI model

Meta's New Model: Muse Spark

Meta has recently unveiled Muse Spark, their latest artificial intelligence model, marking their first major announcement since Llama 4 nearly a year ago. Unlike some previous models, Muse Spark is not open weight, meaning it is hosted and accessible only via a private preview API. This API is currently reserved for selected users, but it is possible to test Muse Spark right now on meta.ai, provided you log in with a Facebook or Instagram account.

The performance of Muse Spark, as reported by Meta, indicates that it is competitive with models like Opus 4.6, Gemini 3.1 Pro, and GPT 5.4 on certain specific benchmarks. However, it is important to note that Muse Spark lags behind on the Terminal-Bench 2.0 benchmark. Meta has acknowledged these shortcomings and claims to continue investing in performance improvements, particularly in the areas of long-term agent systems and coding workflows.

Usage Modes and Integrated Tools

On the meta.ai platform, Muse Spark is offered in two distinct usage modes: "Instant" and "Thinking." The "Instant" mode is designed for quick responses, while the "Thinking" mode allows for a slightly longer reflection time. Meta has also announced the development of a "Contemplation" mode, which promises to significantly extend reasoning time, approaching the capabilities of models like Gemini Deep Think or GPT-5.4 Pro.

The Muse Spark interface is enhanced with several integrated tools that expand its capabilities:

• Search and Navigation: The browser.search tool allows for web searches via a search engine whose name has not been disclosed. With browser.open, it is possible to load the full page of a search result, and browser.find enables pattern matching on the content of the returned pages.

• Meta Content Search: The meta_1p.content_search tool offers semantic search across posts on Instagram, Threads, and Facebook. This feature is limited to posts visible to the user and created since January 1, 2025.

• Image Generation: With media.image_gen, users can generate images from prompts. The generated images are stored in a sandbox and can be retrieved via a CDN URL. Two generation modes are available: "artistic" and "realistic."

• Python Code Execution: The container.python_execution tool allows for the execution of Python code in a remote sandbox environment, using Python 3.9 and several useful libraries.

• Web Artifact Creation: With container.create_web_artifact, it is possible to create HTML+JavaScript files in a secure container, which can be used as interactive iframes.

• Image Analysis: The container.visual_grounding tool can analyze an image, identify and label objects, locate specific regions, or count objects.

Image Analysis and Generation

Muse Spark demonstrates an impressive ability to generate and analyze images. For example, a request to create an image of a raccoon sitting on a trash can, wearing trash as a hat, produced a humorous image. This image was then analyzed using the available Python tools, showcasing Muse Spark's image analysis accuracy.

Using the container.visual_grounding tool, each element of the image was identified with remarkable precision. The results included details such as:

• Raccoon: a large box covering 62% of the image width.
• Coffee Cup: positioned at the top of the image, measuring 158 pixels in height.
• Banana Peel: overlapping with the base of the cup.
• Newspaper: covering the raccoon's left ear.
• Trash Can Lid: extending almost the entire width of the image at the bottom.

The tool even has the capability to count the raccoon's whiskers, illustrating its potential for complex image analysis tasks.

Conclusion

With the launch of Muse Spark, Meta continues to advance in the field of artificial intelligence, offering advanced tools for interaction and image analysis. While performance challenges remain, this model represents a significant step forward for Meta in the AI race.