Meta just dropped Muse Spark, their first model since Llama 4 about a year ago, and it’s a weird one. After spending years building trust with developers through open weights releases, they’ve gone fully hosted. No weights, no local runs, just a “private API preview” and access through meta.ai if you’re willing to log in with Facebook or Instagram.
The benchmarks look competitive with the big players like Opus 4.6, Gemini 3.1 Pro, and GPT 5.4 on selected tests. Meta admits they’re behind on Terminal-Bench 2.0 and still working on “long-horizon agentic systems and coding workflows.” At least they’re honest about the gaps.
What caught my attention wasn’t the model itself but the absolutely massive tool collection they’ve wired into the chat interface. We’re talking 16 different tools, and Meta’s chatbot will actually tell you about them if you ask. No jailbreaking required, which is refreshing compared to the usual corporate paranoia about exposing system internals.
The Tool Arsenal
The browser tools are standard fare at this point. You get browser.search for web searches, browser.open to load full pages, and browser.find for pattern matching. Nothing groundbreaking, but table stakes for any artificial intelligence assistant trying to be useful.
What’s more interesting is meta_1p.content_search. This thing can do semantic search across Instagram, Threads, and Facebook posts, but only for content you have access to and only since January 1st, 2025. The parameters are wild: author_ids, key_celebrities, commented_by_user_ids, liked_by_user_ids. I can see why they’re keeping this one locked down to your own content permissions.
There’s also meta_1p.meta_catalog_search for shopping, which makes sense given Meta’s commerce ambitions. I haven’t tested this one because I’m not trying to buy anything from Facebook, but the integration is there.
Image generation comes through media.image_gen with “artistic” and “realistic” modes. My guess is this is powered by their Emu model or something descended from it. Same pattern as ChatGPT and Gemini where image generation happens via tool call rather than being baked into the model itself.
Code Interpreter, Finally
Here’s where it gets fun. Meta AI has container.python_execution, which is basically Code Interpreter. Python 3.9.25 with the usual suspects: pandas, numpy, matplotlib, plotly, scikit-learn, PyMuPDF, Pillow, OpenCV. Python 3.9 is technically EOL now, but the library selection is solid enough for most use cases.
Files persist at /mnt/data/, and you get SQLite 3.34.1 from January 2021. Not cutting edge, but functional.
The really clever bit is how all these tools compose. I generated an image of a raccoon wearing trash as a hat (because why not), and since image_gen saves images to the sandbox, I could immediately run OpenCV analysis on it. The system automatically chained the tools together without me having to manually pass file paths around.
container.create_web_artifact is their version of Claude Artifacts. You can create HTML and JavaScript files that get served as secure sandboxed iframes. When I ran my pelican SVG test, the “Instant” model spat out raw SVG while the “Thinking” model wrapped it in an HTML shell with some unused Playables SDK v1.0.0 JavaScript libraries. Different reasoning modes, different output preferences apparently.
Visual Grounding Is the Star
The most interesting tool is container.visual_grounding. This thing can analyze images, identify and label objects, locate regions, or count items. It returns results in three formats: bbox (bounding boxes), point (center coordinates), or count (how many of something).
I asked it to analyze that raccoon image and it pinpointed every piece of the outfit with pixel-level precision. Face inside raccoon, eyes inside face, hat pieces overlapping but staying above the face box. Real object localization, not guessing.
The bbox mode draws proper nested bounding boxes. When I switched to count mode and asked about whiskers, it returned actual counts. Meta AI can count a raccoon’s whiskers out of the box. Is this useful? Probably not for most people, but it shows the level of detail these vision tools can handle.
It doesn’t do segmentation masks directly, but since you have Python in the container, you can fake it with OpenCV. The model suggested generating alpha masks to cut out the raccoon or individual hat pieces. Ten seconds of processing for a transparent PNG.
The Rest of the Tools
There’s container.download_meta_1p_media for pulling in content from Instagram, Facebook, and Threads posts directly into the sandbox. Combined with Code Interpreter, you could do some interesting analysis of your own social media content.
container.file_search does semantic search over uploaded files. container.view, container.insert, and container.str_replace are text editor commands, similar to what Claude uses for file manipulation.
subagents.spawn_agent follows the sub-agent pattern we’re seeing everywhere now. Spawn an independent agent for research or analysis, get back a final text response.
third_party.link_third_party_account can connect Google Calendar, Outlook Calendar, Gmail, or Outlook. I haven’t tested this because I don’t want Meta having OAuth tokens to my email, but the capability exists.
But Where Are the Weights?
Here’s the problem. After years of Llama releases that developers could download and run locally, Meta has pivoted to hosted-only. The API is currently limited to “select partners” and there’s no timeline for general availability.
Jack Wu from Meta says “bigger models are already in development with infrastructure scaling to match” and promises plans to “open-source future versions.” I really hope they follow through on that. Llama 3.1, 3.2, and 3.3 were excellent laptop-scale model families that developers actually used.
The intro blog post claims they can “reach the same capabilities with over an order of magnitude less compute than our previous model, Llama 4 Maverick.” If that efficiency translates to smaller, open-weights releases, we’d have something genuinely useful for local deployment.
Artificial Analysis scored Muse Spark at 52, behind only Gemini 3.1 Pro, GPT-5.4, and Claude Opus 4.6. Last year’s Llama 4 Maverick and Scout scored 18 and 13 respectively. So yes, Meta is back in the frontier model race, but they’re playing by different rules now.
The tool collection is impressive and the composability is well thought out. But I can’t really evaluate this model properly without API access. What developers need isn’t a fancy chat interface with lots of tools, it’s the ability to build their own systems on top of reliable, affordable inference. Until Meta opens up access or releases weights, Muse Spark is just another demo that most of us can’t actually use for anything serious.