Meta just dropped Muse Spark, their first major model release in a year. The benchmarks show it competitive with Claude Opus 4.6 and GPT 5.4. But thats not the interesting part.
Whats interesting is what Simon Willison discovered when he started poking around the meta.ai interface. He asked a simple question: what tools do you have access to?
The answer revealed 16 tools. And Meta didnt hide them.
The Tool Stack Nobody Mentioned
Heres what Meta quietly shipped:
Browser tools. browser.search, browser.open, browser.find. Web search through an undisclosed engine, page loading, and pattern matching against content. Basic but essential.
Meta content search. meta_1p.content_search can search Instagram, Threads, and Facebook posts semantically—but only for content the user can access, created since 2025-01-01. Parameters include author_ids, key_celebrities, commented_by_user_ids, liked_by_user_ids. Thats a lot of filtering power.
Code Interpreter. container.python_execution runs Python 3.9 in a sandbox with pandas, numpy, matplotlib, plotly, scikit-learn, PyMuPDF, Pillow, OpenCV. Files persist at /mnt/data/. Sound familiar? Its the same pattern ChatGPT and Claude use.
Web artifacts. container.create_web_artifact creates HTML+JavaScript files that render as sandboxed iframes. Set kind to html for apps or svg for graphics.
Visual grounding. This one is fascinating. container.visual_grounding analyzes images, identifies objects, and returns bounding boxes, points, or counts. Its Segment Anything as a tool—ask it to count whiskers on a raccoon and it outputs coordinates for each one.
Subagent spawning. subagents.spawn_agent delegates tasks to independent sub-agents. The pattern Simon documented months ago is now a built-in tool.
Why This Matters
The model itself is fine. Artificial Analysis scores it at 52, behind only Gemini 3.1 Pro, GPT-5.4, and Claude Opus 4.6. Meta claims over an order of magnitude less compute than Llama 4 Maverick.
But the real story is the convergence. Every major AI company is arriving at the same tool architecture:
- Python execution sandbox
- Web artifact rendering
- File manipulation primitives (view, insert, str_replace)
- Visual analysis grounded in the sandbox
- Subagent delegation
Metas implementation includes their own twist: tight integration with their social graph. Thats a moat. Thats data Claude and GPT cant access.
The Open Weights Question
Alexandr Wang hinted at open-sourcing future versions. Meta pioneered open weights with Llama. Then went closed with Llama 4. Now maybe back again?
If they release Muse Spark weights, the tool harness becomes a reference implementation. Developers could replicate the meta.ai experience locally.
But for now, its hosted only. Private API preview for select users. The tools work—but youre renting them, not owning them.
The Takeaway
The model race gets attention. The tool race matters more.
Metas 16-tool harness is sophisticated. Code Interpreter + visual grounding + subagent spawning + social graph search. Thats a productivity stack, not a chatbot.
Claude has similar capabilities. GPT has similar capabilities. Gemini has similar capabilities. Were not comparing models anymore. Were comparing tool ecosystems.
And the companies building the best tools—not just the smartest models—will win.
Top comments (0)