Kathmandu – Meta has introduced two new AI models under its Llama 4 framework. These models will boost Meta AI Assistant on WhatsApp, Messenger, Instagram, and other platforms.

The models, Llama 4 Scout and Llama 4 Maverick, are available for download on Meta and Hugging Face. Llama 4 Scout is compact and runs on a single NVIDIA H100 GPU. Llama 4 Maverick is designed to compete with OpenAI’s GPT-4o and Google’s Gemini 2.0 Flash.

Key Features

Meta CEO Mark Zuckerberg announced that Meta is also training Llama 4 Behemoth. He claims it will be “the highest-performing base model in the world.”

Llama 4 Scout has a 100 million token context window, making it efficient for handling large data. It outperforms Google’s Gemini 3, Gemini 2.0 Flash-Light, and Mistral 3.1 while running on a single GPU.

Llama 4 Maverick is designed for coding and reasoning tasks. Meta says it outperforms OpenAI’s GPT-4o and Google’s Gemini 2.0 Flash. It is comparable to DeepSeek-V3, a model known for advanced reasoning.

Llama 4 Behemoth is expected to be even more powerful. It will have 288 billion active parameters and 2 trillion total parameters. Meta claims it will surpass GPT-4.5 and Claude Sonnet 3.7 in STEM-related tasks.

Meta is using the Mixture of Experts (MoE) architecture for Llama 4. This means only relevant parts of the model activate for each task, saving resources.

What’s Next?

Meta will share more details at LlamaCon on April 19. The event will showcase Meta’s AI roadmap and future innovations.