Meta unveils 2 new multi-modal AI models: Llama 4 Scout and Maverick | Technology News

Of the two, Llama 4 Scout is a small multimodal model with a 10 million token context length that can run on a single GPU such as the Nvidia H100. With 17 billion parameters across 16 experts, it features a Mixture of Experts (MoE) architecture and is considered the highest-performing model in its class.

In the same Instagram Reel, Zuckerberg described it as a “workhorse” that outperforms GPT-4o and Gemini Flash 2 in all benchmarks. It is also said to be smaller and more efficient than DeepSeek V3, which is another 17-billion-parameter model using 128 experts. Llama 4 Scout can also run on a single host for easy inferencing.

Today is the start of a new era of natively multimodal AI innovation.

Today, we’re introducing the first Llama 4 models: Llama 4 Scout and Llama 4 Maverick — our most advanced models yet and the best in their class for multimodality.

Llama 4 Scout
• 17B-active-parameter model… pic.twitter.com/Z8P3h0MA1P

— AI at Meta (@AIatMeta) April 5, 2025

Llama 4 Reasoning is set to launch next month, while Llama 4 Behemoth — a large language model with two trillion parameters — is currently in training and is expected to be the biggest model yet.

Llama 4 Scout and Llama 4 Maverick are already available for free download via Meta and Hugging Face. However, there may be certain licensing restrictions that limit access to these open-source models.