Meta's last major model release was Llama 4, about a year ago. Their follow-up, Muse Spark, dropped on April 8 - and it makes a quiet but significant departure from Meta's usual open-source positioning: the weights aren't public.
Muse Spark is currently in a private API preview for select developers, but anyone with a Facebook or Instagram login can access it through meta.ai today. Meta claims the model achieves comparable results to Claude Opus 4.6, Gemini 3.1 Pro, and GPT 5.4 on selected benchmarks, though it trails on Terminal-Bench 2.0, a coding and terminal task evaluation. The bigger engineering claim is efficiency - Meta says Muse Spark delivers similar capability with more than ten times less compute than Llama 4 Maverick, their largest previous model.
The model has two current modes - Instant (fast, lower-effort responses) and Thinking (the model reasons step-by-step before answering, similar to how ChatGPT's o-series works) - with a third "Contemplating" mode promised later, analogous to Gemini's Deep Think.
What's Actually Inside meta.ai
The more interesting part isn't the model benchmarks - it's the 16 tools wired into the meta.ai chat interface. Most AI assistants offer web search and maybe a calculator. Meta's list is longer and more specific:
- Web search via an undisclosed provider
- Social search across Instagram, Threads, and Facebook posts from January 2025 onward - content the logged-in user has access to
- Python execution running Python 3.9 with scientific libraries including pandas, numpy, matplotlib, scikit-learn, and OpenCV
- HTML/JavaScript artifact creation with sandboxed preview, similar to Claude's Artifacts feature
- Image generation in artistic or realistic modes, across square, vertical, and landscape formats
- Visual grounding - the model can identify and count objects in photos with bounding-box precision. In one test, it correctly counted 25 pelicans in a single image
- Calendar and email access via linked Google Calendar, Outlook, and Gmail accounts
- Sub-agent spawning - the model can delegate tasks to secondary agents internally
The Closed-Weights Question
For daily users of meta.ai, none of this matters much - you just get a capable assistant with useful tools. But for the developer community that built products on Llama's open weights, Muse Spark is a signal worth watching. Meta has always differentiated itself from OpenAI and Anthropic by releasing model weights publicly. Muse Spark is a hosted-only product, at least for now.
That may be temporary - Meta hasn't ruled out a future open release. But right now, Muse Spark ships more like a commercial product than a research artifact. Given the compute efficiency claims, keeping it proprietary makes financial sense for Meta even if it frustrates the open-source community they've spent years cultivating.
For people who just want an AI assistant that can search their Instagram history, run Python on a dataset, or count objects in a photo, meta.ai is now a more capable daily driver than most people probably realize.