Meta just rolled out a fresh batch of AI models under the Llama 4 name, and they’re already running inside Meta AI on WhatsApp, Instagram, and Messenger.

You've got Scout—the chill, low-lift one that runs on a single GPU; Maverick—the more intense version built to spar with GPT-4o; and Behemoth—which isn’t quite ready yet but sounds massive, with 288 billion active parameters in the works.

Scout’s pulling off some solid numbers, beating out models like Gemini Flash-Lite, all without needing a ton of compute.

Maverick’s more of a scrappy overachiever, competing with the big names while staying efficient.

Then there’s Behemoth, which Meta is hyping as the top-performing base model out there—we’ll see what it’s really made of at LlamaCon on April 29.

Meta’s calling these models “open-source,” but there’s a catch.

If your platform has over 700 million monthly users, you’ve gotta ask Meta first. So yeah, that’s sparked some open-source side-eye.