Meta has just released Llama 4, its newest set of artificial intelligence models, designed to process and generate text, images, audio, and video. The first two publicly available versions—Llama 4 Scout and Llama 4 Maverick—are now open for download, while a more advanced model, Llama 4 Behemoth, remains in development.  

What's in Llama 4?  

The release includes two main models. Llama 4 Scout is a smaller, efficient model that can run on a single high-end GPU, with a large memory capacity for processing long documents or conversations. Llama 4 Maverick is a more complex model that Meta claims performs comparably to leading AI systems like GPT-4o and Gemini 2.0 in some tasks, while using fewer computing resources.  

Both are "multimodal", meaning they can work with different types of data—such as analysing a photo, summarising a video, or answering questions about an audio clip. Meta says these models improve upon earlier versions in reasoning and coding tasks.  

How could it be used?

The technology could enhance AI features in Meta's apps like WhatsApp, Messenger, and Instagram, allowing for more advanced chatbots or tools that interpret media. Developers and businesses may also integrate the models into their own software for tasks like customer support, content moderation, or data analysis.  

Unlike some competitors, Meta is making Llama 4 Scout and Maverick freely available as open-source software, allowing outside developers to modify and build on the technology. The company is still refining Llama 4 Behemoth, a more powerful model intended to assist in training future AI systems.  

Users can test the models now through Meta's website or third-party platforms like Hugging Face.

Meta's push into AI follows industry-wide investments sparked by the success of ChatGPT. Reports from Reuters and The Information suggest Llama 4 faced delays due to performance concerns, particularly in matching rivals' conversational abilities. The company plans to spend heavily on AI infrastructure this year, though it remains to be seen how these models will compete in a rapidly evolving market.