Meta has launched the Llama 4 series with two new AI models: Llama 4 Scout and Llama 4 Maverick. Both feature a multimodal, mixture-of-experts (MoE) architecture and 17 billion active parameters.

Scout uses 16 experts and is optimized for a single Nvidia H100 GPU, offering a 10 million token context window and outperforming several rivals in benchmarks.

Maverick, with 128 experts, focuses on efficiency and cost-effectiveness, delivering strong results in reasoning and coding tasks.

Developers can access both models on platforms like Hugging Face and the Llama website.

Meta also previewed Llama 4 Behemoth, a 288-billion-parameter model still in training. It shows strong performance on STEM-focused benchmarks.

The Llama 4 series was trained on diverse datasets, including text, images, and videos, using techniques like MetaP and FP8 precision to boost quality and efficiency.

Supporting over 200 languages, these models are compatible with platforms like WhatsApp, Messenger, and Instagram Direct. More updates will be shared at LlamaCon on Apr. 29.

Recent Meta developments