With some licensing limitations, Meta introduces llama 4 Scout and Maverick to the market – the first for the company in this frontier.
In an Instagram reel published on Saturday, Mark Zuckerberg introduced Meta’s two new open-source AI models. These are the first of their kind in the Llama 4 family.
These next-generation language models – Llama 4 Scout and Llama 4 Maverick – are said to be a stepping stone for both Meta and the open-source community.
Llama 4 Scout, Zuckerberg describes, is a small multi-modal model running on a single GPU, just like the Nvidia H100. He goes on to label it the “workhorse,” capable of outperforming Gemini Flash 2 and ChatGPT 4o across all benchmarks.
This model leverages 17 billion parameters across 16 experts and can run on a single host for easy inference. Even more so, Llama 4 Scout entails MoE architecture, establishing it as the highest-performing model in its class.
Meanwhile, Llama 4 Maverick is a 17B active parameter model with 128 experts. It has one of the best grounding abilities, helping align the users’ prompts with relevant images.
Beyond these two models, Meta is set to launch two more in the same family – Behemoth and Reasoning. Both language models will entail 1 trillion parameters. There’s no single launch date as the models continue to be under training.
Speculations are that this could be the tech powerhouse’s enormous model yet.
It’s intriguing to witness the steps Meta is taking toward concretizing the AI economy. Once thought of as a mere trend, it’s become the nucleus of our daily lives. And as new products in the AI field keep on appearing, we might be led to believe, albeit with doubt, that we haven’t tapped even half its potential yet.