What it does
Llama is a family of models you can download from this site. The newest generation, Llama 4, is described as natively multimodal: Maverick handles image and text understanding, while Scout is aimed at long-document analysis, with single H100 GPU efficiency and a 10M-token context window. The Llama 3 line includes 3.1 (8B, 70B and 405B), 3.2 (lightweight 1B and 3B models plus multimodal 11B and 90B) and 3.3 (70B). The site also hosts guides on prompt engineering, fine-tuning, vision capabilities, quantization, distillation and evaluations, along with a cookbook, videos, deployment and integration guides, case studies and a developer use guide. A benchmark table compares Llama 4 Maverick and Scout on reasoning, coding, multimodal, multilingual and long-context tasks. Beyond model downloads, the navigation lists a Llama API (with a waitlist and login), the Llama Stack, and Llama Protections safety tooling including a Defenders Program. Links lead to Hugging Face and GitHub for model files and code.
Capabilities
- Downloadable Llama 4 (Scout, Maverick) and Llama 3.1, 3.2 and 3.3 models
- Native text and image understanding in the Llama 4 models
- Small 1B and 3B Llama 3.2 models positioned for edge use
- How-to guides for fine-tuning, quantization, distillation and evaluations
- Llama Stack, a Llama API waitlist and Llama Protections safety tools
When to use it
- You want open models to fine-tune, distill or deploy on your own infrastructure
- You need a model that reasons over images with text, or over very long documents
- You are looking for model cards, prompt formats and deployment guides for Llama
