Llama 2

Native multimodal models (text+vision) with early fusion and joint pre‑training on text, image, and video data

Industry‑leading long context: Llama 4 Scout supports up to 10 million tokens (trained at 256K) for strong length generalization

Mixture‑of‑Experts architecture: Scout (17B active/109B total, 16 experts) and Maverick (17B active/400B total, 128 experts)

Open‑weight and open‑source availability with 300M+ cumulative downloads and a permissive license for model output use

Extensive multilingual coverage: pre‑training across 200 languages with 10x more multilingual tokens than Llama 3

Edge/on‑device deployment: Llama 3.2 1B and 3B models run locally for private, low‑latency experiences

State‑of‑the‑art performance on STEM, coding, reasoning, multilingual, long‑context, and image benchmarks

Efficient training at scale: MetaP for stable hyper‑parameters, >30T token data mix, FP8 training on 32K GPUs achieving ~390 TFLOPs/GPU

Advanced tool use and reasoning, with small models outperforming peers on instruction following and tool-calling tasks

Broad range of model sizes from 1B to 405B parameters to fit mobile, server, and frontier workloads