Meta has officially unveiled its latest Llama 4 AI models, positioning itself as a serious contender in the competitive AI landscape. With the launch of two distinct models, one that is lightweight enough to operate on a single GPU and another that is considerably larger, Meta aims to challenge existing heavyweights like OpenAI, Google, and DeepSeek. The two new Llama 4 models will enhance the AI functionality across several of Meta’s platforms, including Instagram, Facebook Messenger, and WhatsApp.

The first model, named Llama 4 Scout, features 17 billion parameters and utilizes 16 experts, resulting in a total of 109 billion active parameters. One of its standout features is a remarkably large context window of 10 million tokens, a substantial upgrade from the impressive 1 million-token context window introduced by Gemini. Despite its advanced capabilities, Llama 4 Scout remains compact enough to be utilized on a single Nvidia H100 GPU, making it more suitable for personal applications and modest tasks.

In contrast, Llama 4 Maverick serves as the larger counterpart, designed for heavy industry use. It boasts the same 17 billion parameters but expands its functionality with 128 experts, ultimately reaching 400 billion total parameters. While the context window here is limited to 1 million tokens, its robust specifications cater more to business needs and larger projects.

Looking ahead, Meta is also preparing to introduce Llama 4 Behemoth, its most powerful model, expected to feature 288 parameters and 2 trillion active parameters. Currently in preview, the launch date remains unknown, but details are anticipated at LlamaCon on April 29th. This upcoming model promises to position Meta at the forefront of AI innovation.