Part 2/8:
Llama 4 introduces three major variants, two of which are currently available for immediate download while the Behemoth model is still in development. Each version is built on mixture of expert models, allowing different parts of the model to excel in specific tasks. This strategic design enhances the overall efficiency and performance of the AI.
- Llama 4 Scout: The entry-level model with 109 billion total parameters and 17 billion active parameters, featuring 16 experts. It operates with a remarkable 10 million token context length, surpassing competitors like Gemini, which only provides 2 million tokens.