Meta has released a new series of AI models, Llama 4

Meta has announced a new series of Llama 4 AI models, which includes two new models: Llama 4 Scout and Llama 4 Maverick. These models are already used by Meta AI on the website, as well as in WhatsApp, Messenger and Instagram.

The Llama 4 Scout is a smaller model that can run on a single NVIDIA H100 chip and outperforms the Google Gemini 3 and Gemini 2.0 Flash-Lite in several benchmarks. The Llama 4 Maverick, on the other hand, has similar performance to the GPT-4o and Google Gemini 2.0 Flash, showing good results in coding and logic tasks, while using fewer parameters.

The new series will also feature the Llama 4 Behemoth model, which has 288 billion active parameters and 2 trillion total parameters. It is still in the training process, but promises to be the most powerful model yet. Meta claims that Behemoth will outperform competitors such as GPT-4.5 and Claude Sonnet 3.7 on several STEM benchmarks.

To develop Llama 4, Meta used a "mixing of experts" (MoE) architecture, which allows for efficient use of resources by activating only those parts of the model that are necessary to perform specific tasks.

The company plans to share its future plans for AI models at the LlamaCon conference, which will take place on April 29.