Meta Releases Llama 4, a Powerful New Lineup of AI Models

Meta has launched its new Llama 4 series of AI models, which are now being used in the Meta AI assistant. This assistant now works on WhatsApp, Messenger, Instagram and websites. This time Meta has introduced three models – Llama 4 Scout, Llama 4 Maverick and Llama 4 Behemoth. Scout and Maverick have been made available to general developers, while Behemoth is currently in the training phase and will be released later. These models can be downloaded from platforms like Meta and Hugging Face.

Scout is the lightest model, which can be run on a single Nvidia H100 GPU. Its special feature is its large context window with 10 million tokens, which allows it to understand and process long documents, code, and complex information well. It is perfect for small but effective tasks and according to Meta, it performs better than models like Google’s Gemma 3 and Gemini 2.0 Flash-Lite. Maverick, on the other hand, is a mid-level powerful model pitted against larger models like GPT-4o and Gemini 2.0 Flash. It gives better results in many areas like coding, logical reasoning, multi-language support and image processing. The special thing is that Maverick has 400 billion total parameters, but only 17 billion parameters are active at a time, which increases its speed and efficiency.

Scout has 109 billion total parameters and 16 experts, while Maverick has 128 experts. Both models use Mixture of Experts (MoE) architecture. This technology works in such a way that different “expert” models are activated as per the need for each task, which saves computing resources and the model works quickly and smartly.

The third model Behemoth has been described by Meta as the most powerful base model till date. It will have 2 trillion total parameters and 288 billion active parameters. This model has not been made public yet, but Meta claims that it will perform better than AI models like GPT-4.5, Claude 3.7 Sonnet and Gemini 2.0 Pro, especially in subjects like STEM i.e. mathematics and science.

Meta said that Scout and Maverick models are being used in Meta AI in 40 countries of the world, but its multimodal features (such as understanding image input with text) are currently available only in America and English language. Meta also said that this time their AI models do less ‘censorship’ than before. That is, now these models are more open in answering political and social controversial questions. Meta says that they have tuned Llama 4 in such a way that it can answer in a more balanced and non-judgmental manner.

Although Meta has called these models “open-source”, there are still some limitations in its license. For example, companies with more than 700 million monthly active users must obtain separate permission from Meta. Additionally, individuals or companies located in the European Union cannot use or distribute it, possibly due to EU data privacy and AI regulations.

Meta CEO Mark Zuckerberg has said that the Behemoth model will be the most powerful AI model in the world in the future. Along with this, Meta will share its upcoming plans at an event called LlamaCon on April 29, where it will reveal its next AI products and research.

Related Post :- Don’t Miss It: How to Watch Meta’s First-Ever AI Developer Event – LlamaCon