Meta · Llama

Llama 4

Llama 4 is Meta’s current open-weight generation. Public cards center on Llama 4 Scout and Llama 4 Maverick (Mixture-of-Experts style naming on Meta and GitHub herd tables). Meta’s Llama 4 developer page frames them as optimized for deployment and efficiency.

Status: Current, open weights

Training

Parameter and expert layouts on Meta and GitHub herd notes include Scout-17B-16E and Maverick-17B-128E. Full FLOPs and data mix are not published unless a Meta model card states them. License is the Llama community license (not Apache). Read exact license text from Meta’s page.

Inference

Open weights via Meta Llama download flow and Hugging Face meta-llama repos. Multimodal image reasoning scores appear on Meta’s Llama 4 page. Serving silicon for end users is the buyer’s hardware or cloud.

Context

GitHub llama-models README associates Scout with very long context (10M note) and Maverick with 1M. Confirm on live Meta model cards before treating those figures as final.

Methods

Open-weight Llama post-training stack. Exact RL recipe for Scout and Maverick is not published unless a model card says so.

Benchmarking

Use only scores printed on the Llama 4 developer page or official model cards. Examples Meta lists there include MMMU, MathVista, ChartQA, DocVQA, and others.

Hardware

Meta has not published per-SKU training silicon for Llama 4 on the pages cited here. Buyers run weights on their own GPUs or hosts.