Zuckerberg said: “Our goal is to build the world’s leading AI, open source it, and make it universally accessible so that everyone in the world benefits.”

He unveiled Llama 4 Scout and Llama 4 Maverick — with two more, including a behemoth, still to come.

“I’ve said for a while that I think open-source AI will become the leading model,” he added. “And with Llama 4, this is starting to happen.”

Scout has 17 billion parameters, 16 experts, and a “nearly infinite” 10 million-token context. It’s natively multimodal, runs on a single GPU (Meta says an NVIDIA H100), and is being touted as the top small model in its class.

Maverick has 17 billion parameters, but with 128 experts, the workhorse outperforms GPT -4 and Gemini Flash 2 on all benchmarks while still being more efficient than DeepSeek V3. “This thing is a beast,” Zuckerberg said, not even pretending to play it down.

The heavyweight arriving later is Llama 4 Behemoth — tipping the scale at more than two trillion parameters, with 288 billion active. Meta claims it outperforms GPT-4.5, Claude Sonnet 3.7, and Gemini 2.0 Pro on STEM benchmarks and will serve as a “teacher” model for future AI builds. 

Zuckerberg said, “I’m not aware of anyone training a larger model out there.” It's already the highest-performing base model in the world, and it's not even finished training yet.

There is also a fourth model in the pipeline, called Llama 4 Reasoning, with more to be revealed at “LlamaCon” on April 29.

Meta says you can now test-drive Scout and Maverick through Meta AI on WhatsApp, Messenger, or Instagram Direct or grab them from llama.com or Hugging Face. 

Meta’s bog described it as “the beginning of a new era of natively multimodal AI innovation”. It claimed, “Openness drives innovation — and is good for developers, good for Meta, and good for the world.”