The Llama 4 herd: natively multimodal AI

Summary

Meta's launch post for Llama 4, its first mixture-of-experts and natively multimodal Llama models. It released two open-weight models, Scout and Maverick, and previewed Behemoth, a much larger teacher model that was still in training.

Key points

• Scout: 17B active and 109B total parameters, 16 experts, 10M-token context window.
• Maverick: 17B active and 400B total parameters, 128 experts.
• Behemoth (preview, not released then): 288B active, about 2T total, 16 experts.
• Text, image, and video input.
• Weights on llama.com and Hugging Face under Meta's custom Llama 4 licence, which is not an OSI open-source licence and adds usage conditions.

Why it matters

Brought MoE and very long context to the Llama line. As of October 2026 it is the most recent Llama release recorded in this library.