Llama 4, Meta releases two multimodal artificial intelligence models

And then there is Llama 4 Maverick designed as a general assistant and for creative writing. The same size as its little brother with 128 experts and a total of 400 billion parameters, it is the best multimodal model in its class, surpassing GPT-4o and Gemini 2.0 Flash in a wide range of benchmarks, achieving results comparable to DeepSeek v3 on reasoning and coding, with less than half the parameters active.

In addition, for the new models, Meta says it is adopting a mixed-of-experts (MoE) architecture, an approach that aims to save resources by using only the parts of the model that are needed to perform a given task broken down into smaller tasks: "We are introducing Llama 4 Scout and Llama 4 Maverick, the first open-weight native multimodal models with unprecedented context length support and our first built using a mixed-of-experts (MoE) architecture. We are also previewing Llama 4 Behemoth, one of the world's smartest LLMs and our most powerful to date to serve as a teacher for our new models,' writes the Meta team in the official announcement blogpost.

Llama 4 Behemoth, in fact, is a multi-modal model still in the training phase that, with 288 billion active parameters, 16 experts and almost two trillion total parameters, will act as a 'teacher' for the smaller Llama 4 models, guaranteeing high-level performance superior to GPT-4.5, Claude Sonnet 3.7 and Gemini 2.0 Pro according to several STEM benchmarks.

As announced by the Menlo Park-based company, the capabilities of Llama 4 Scout and Llama 4 Maverick can already be tried out in Meta AI on WhatsApp, Messenger, Instagram Direct and the Meta.AI website, and are also available for download on llama.com and Hugging Face. Availability at the company's partners is expected soon so that everyone can experience the newly released models right away. Further details on the models and related products will be revealed during the LlamaCon conference scheduled for 29 April.