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

Simon Willison is arguably the most insightful "archaeologist" in the AI blogosphere. While others obsess over benchmark scores when a new model drops, he's busy asking the AI, "What tools do you have at your disposal?" This seemingly simple question unearthed 16 hidden tools powering Meta AI's Muse Spark.

Muse Spark itself is the star of the show: Meta's latest model, arriving almost a year after Llama 4. It performs comparably to Opus 4.6 and Gemini 3.1 Pro on leaderboards, but Meta openly admits it still lags in "long-horizon agent tasks" and "coding workflows." This honesty is refreshing – a stark contrast to some vendors who cherry-pick only the most flattering benchmarks.

But what I find truly compelling are the 16 tools Simon discovered.

The Browser Toolkit: browser.search, browser.open, browser.find. This means Muse Spark has genuine web browsing capabilities – it doesn't just suggest you search for something, it can actually perform the search, load the page, and locate specific content within it. This is foundational infrastructure for building true agent capabilities.

Meta Content Search: This is the most distinctive tool. meta_1p.content_search can retrieve content across Instagram, Threads, and Facebook, with filters for author, celebrity, commenter, and liker. This is a very "Meta" tool – leveraging their unique data assets from their social platforms. In other words, if you ask, "That post I made on Instagram last week," it can actually find it for you.

Python Execution: container.python_execution, Simon's personal favorite. A complete Python 3.9 environment, pre-loaded with libraries like pandas, numpy, matplotlib, and scikit-learn. This means you can perform data analysis and create visualizations directly within the conversation, without switching to a Jupyter Notebook.

Here's an interesting engineering choice: Instant mode outputs SVG code directly, while Thinking mode wraps the output in an HTML shell, complete with some "unused Playables SDK v1.0.0" libraries. Simon jokingly says, "This makes me curious" – and rightly so. Different reasoning modes might be calling different rendering paths.

Simon also praised Meta for allowing users to directly ask "What tools do you have?" instead of forcing them to use jailbreaking techniques to uncover them. This transparency is developer-friendly and suggests that Meta, at least in the current version, is committed to a degree of openness.

What's the takeaway from all this? AI chat interfaces are evolving from "talking tools" into "collections of tools." When a model can use a browser, execute code, and search social media, the chat box becomes more than just a Q&A interface – it becomes an operational portal. This aligns with Anthropic's MCP protocol and OpenAI's agent tools, but each company is taking a different path.

For the average user, this means you'll soon be able to accomplish more within meta.ai: searching for posts, writing code, analyzing data, all within a single conversation. For developers, this provides a window into the engineering practices of a major tech company – how they organize tool calls, how they handle tool distribution across different reasoning modes, and how they balance feature richness with system stability.

Muse Spark itself might not be revolutionary, but the tool ecosystem behind it reveals a more fundamental trend: the next AI battleground isn't about model parameters, but about "who can enable AI to do more."