Meta announced the release of its new AI models today, dubbed the Llama 4 herd. The company introduced two flagship models, Llama 4 Scout and Llama 4 Maverick, alongside a preview of the still-training Llama 4 Behemoth.
Llama 4 Scout, a 17 billion active parameter model with 16 experts, is designed to fit on a single NVIDIA H100 GPU using Int4 quantization. Meta claims it outperforms all previous Llama models and similarly sized competitors like Gemma 3, Gemini 2.0 Flash-Lite, and Mistral 3.1 across widely reported benchmarks. It boasts an industry-leading context window of 10 million tokens, enabling tasks such as multi-document summarization and reasoning over large codebases.
Llama 4 Maverick, also featuring 17 billion active parameters but with 128 experts and 400 billion total parameters, is designed for top-tier multimodal performance. Meta says it surpasses GPT-4o and Gemini 2.0 Flash on several benchmarks, while achieving results comparable to the much larger DeepSeek v3 in reasoning and coding. Despite its scale, it runs on a single NVIDIA H100 host. An experimental chat version of Maverick has achieved an ELO score of 1417 on LMArena.
Powering these models is Llama 4 Behemoth, a 288 billion active parameter teacher model with 16 experts and nearly two trillion total parameters. Though still in training, Meta reports it outperforms GPT-4.5, Claude Sonnet 3.7, and Gemini 2.0 Pro on STEM-focused benchmarks like MATH-500 and GPQA Diamond. Behemoth plays a key role in distilling knowledge to Scout and Maverick, though it is not yet available for public release.
Both Scout and Maverick employ a mixture-of-experts (MoE) architecture — a first for the Llama series — activating only a subset of total parameters per token to improve efficiency. Scout has 109 billion total parameters, while Maverick scales to 400 billion. The models offer native multimodality with early fusion of text and vision tokens, backed by an enhanced MetaCLIP-based vision encoder.
Developers can download Llama 4 Scout and Maverick starting today, April 5, 2025, from llama.com and Hugging Face. Meta is also rolling out access via partners in the coming days. Users can try Meta AI powered by Llama 4 on WhatsApp, Messenger, Instagram Direct, and the Meta.AI website. More details, including technical insights and future plans for the Behemoth model, will be shared at LlamaCon on April 29.
Hit the link below for the full announcement…
Leave a Reply