Meta's New Model Has 16 Tools. Here's What They Do.
Meta just released Muse Spark—their first model since Llama 4 almost exactly a year ago. It's competitive with GPT-5.4, Gemini 3.1 Pro, and Opus 4.6 on benchmarks. But that's not the interesting part.
The interesting part is what's running under the hood at meta.ai.
Simon Willison poked around and got the model to dump its entire tool catalog—16 tools with full parameter schemas. Meta didn't hide them. No jailbreaks required. Just ask.
Here's what Meta built that everyone else missed.
Code Interpreter, But Better
Python 3.9 sandbox with pandas, numpy, matplotlib, scikit-learn, OpenCV, Pillow. Files persist at /mnt/data/.
The killer feature: you can generate an image with media.image_gen, then immediately analyze it with OpenCV in the same container. Generate a raccoon, count its whiskers. Works.
Visual Grounding Baked In
container.visual_grounding takes an image path and object names, returns bounding boxes, point coordinates, or counts.
This is Segment Anything integrated directly into the chat harness. No separate API call. No context switching. Ask it to locate every piece of a raccoon's trash-hat outfit—it returns pixel-accurate boxes.
[
{"object_name": "raccoon whisker", "count": 12},
{"object_name": "raccoon paw claw", "count": 8},
{"object_name": "trash item on head", "count": 3}
]
Sub-Agent Spawning
subagents.spawn_agent lets Muse Spark delegate research or analysis to independent sub-agents.
This is the sub-agent-as-tool pattern that Claude Code uses—now exposed as a first-class capability in a consumer chat interface.
Meta Content Search
meta_1p.content_search does semantic search across Instagram, Threads, and Facebook posts you have access to.
Filter by:
- Author IDs
- Celebrity mentions
- Comments
- Likes
- Posts since 2025-01-01
This is the moat. No other frontier model has this. ChatGPT can't search your Instagram. Gemini can't see your Threads. Muse Spark can.
Third-Party Account Linking
third_party.link_third_party_account initiates linking for:
- Google Calendar
- Outlook Calendar
- Gmail
- Outlook
The plumbing for the "AI assistant that manages your life" that every company is chasing.
The Full Tool Catalog
-
browser.search- Web search -
browser.open- Load full pages -
browser.find- Pattern match in pages -
meta_1p.content_search- Instagram/Threads/FB search -
meta_1p.meta_catalog_search- Product catalog -
media.image_gen- Image generation (artistic/realistic) -
container.python_execution- Code Interpreter -
container.create_web_artifact- HTML/SVG rendering -
container.download_meta_1p_media- Pull Meta content into sandbox -
container.file_search- Search uploaded files -
container.view/container.insert/container.str_replace- File editing -
container.visual_grounding- Segment Anything -
subagents.spawn_agent- Delegate to sub-agents -
third_party.link_third_party_account- Calendar/email linking
The Pattern Is Clear
Meta isn't just shipping a model. They're shipping a harness—a full agent development environment disguised as a chat interface.
The tools aren't plugins. They're baked in.
And here's the thing: the efficiency claims are absurd. Meta says Muse Spark reaches Llama 4 Maverick capabilities with over an order of magnitude less compute. If true, this changes the economics of frontier inference.
Open Weights?
Alexandr Wang says "plans to open-source future versions."
Not this one. Not yet.
But the efficiency gains suggest laptop-scale models might be coming back to the open ecosystem.
The Real Test
It's not meta.ai. It's what developers can build with the API.
Private preview only for now.
But the tool catalog tells you exactly where Meta is going: not just model provider, but platform for agent work.
The model is the easy part.
The tools are the moat.
Top comments (1)
This is the real shift, not better models, but better systems.
The tool layer is the moat.
Once AI can execute, not just respond, everything changes.