Llama 4

Llama 4 is a language model released on April 5, 2025, featuring a 10 million–token context window. It comes in three versions—Scout, Maverick, and Behemoth—aimed at different workloads, from everyday use to large-scale research.

What Llama 4 can do

Built using a mixture-of-experts (MoE) approach and FP8 training, Llama 4 is designed for fast processing of large inputs. It supports early multimodality for working with more than just text.

• Generate and analyze text with up to 10M tokens of context
• Review very long documents (e.g., books up to ~5,000 pages)
• Write and analyze code
• Process images and audio via early multimodal capabilities
• Support 20 languages (claimed translation accuracy up to 95%)

How to use Llama 4

Llama 4 is available via the official website as a web app and as downloadable models for local use. For local runs, a GPU with 16 GB+ is recommended. In Russia and some other countries, network workarounds may be required due to regional restrictions.

• Sign up on the Llama website
• Download the Scout model (4 GB)
• Install Python 3.10+
• Run the script from a terminal
• Enter a prompt in the interface

The source states Llama 4 is free to use without limits and is open source, allowing code modifications.

*Meta is banned in the Russian Federation.