Meta's new model is Muse Spark, and meta.ai chat has some interesting tools

2026-04-08 · Source: Simon Willison's Weblog · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Emerging Technologies & Innovation · Depth: Intermediate, long

Summary

Meta has released Muse Spark, its first model since Llama 4, available through a private API preview and on meta.ai (requiring Facebook or Instagram login). Benchmarks indicate Muse Spark is competitive with Opus 4.6, Gemini 3.1 Pro, and GPT 5.4, though it lags on Terminal-Bench 2.0. The meta.ai interface offers "Instant" and "Thinking" modes, with a "Contemplating" mode planned for longer reasoning. The platform reveals 16 integrated tools, including `browser.search`, `meta_1p.content_search` for Meta content, `meta_1p.meta_catalog_search` for products, `media.image_gen` for image creation, and `container.python_execution` for sandboxed Python 3.9 code. Notably, `container.visual_grounding` performs image analysis, identifying, locating, and counting objects in images, demonstrated by counting raccoon whiskers and pelicans. The model also supports `subagents.spawn_agent` and `third_party.link_third_party_account` for services like Google Calendar. Future open-sourcing of models is anticipated.

Key takeaway

For Machine Learning Engineers evaluating new LLM platforms, Meta's Muse Spark offers a robust tool-integrated environment, including a Python sandbox and advanced visual grounding capabilities. You should explore its `container.python_execution` and `container.visual_grounding` tools on meta.ai to understand its potential for complex multimodal workflows and agentic system development, especially given the promise of future open-source releases.

Key insights

Meta's Muse Spark model integrates advanced tooling for multimodal interaction and sandboxed code execution.

Principles

• Tool integration enhances LLM capabilities.
• Sandboxed environments enable secure code execution.
• Multimodal models can perform detailed image analysis.

Method

Meta AI's chat harness exposes tools via direct queries, allowing users to inspect tool names, parameters, and descriptions for enhanced interaction and understanding of agent capabilities.

In practice

• Use `container.python_execution` for data analysis.
• Employ `media.image_gen` for creative asset generation.
• Utilize `container.visual_grounding` for object detection.

Topics

• Muse Spark
• Meta AI Platform
• Frontier AI Models
• AI Tool Integration
• Python Code Interpreter

Code references

• facebookresearch/segment-anything

Best for: Machine Learning Engineer, Computer Vision Engineer, CTO, AI Engineer, AI Scientist, Director of AI/ML

Related on AIssential

• [AINews] Meta Superintelligence Labs announces Muse Spark, first frontier model on their completely new stack — Meta Superintelligence Labs launched Muse Spark, its first frontier model on a new stack, with bigger models in development.
• Meta AI Muse Spark IS INCREDIBLE! Powerful Coding & Multimodal Model! (Fully Tested) — Muse Spark is Meta's new multimodal AI, excelling in visual reasoning, coding, and agentic workflows.
• Goodbye, Llama? Meta launches new proprietary AI model Muse Spark — first since Superintelligence Labs' formation — Muse Spark marks Meta's return to frontier AI, emphasizing multimodal reasoning and efficiency over its prior open-source Llama strategy.
• Meta launches Muse Spark as first major AI model from new lab — Meta's Muse Spark aims for competitive AI performance through efficiency and strategic infrastructure investment.
• New ChatGPT Apps Actually Work & More AI Use Cases — AI platforms are rapidly integrating third-party applications and specialized tools to enhance functionality and user experience.
• All of AI's New Models and Tools — AI development is rapidly advancing across multimodal models, open-source frontier models, and agentic platforms.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Simon Willison's Weblog.