news
The corporate launched the primary two fashions from its multimodal Llama 4 suite: Llama 4 Scout and Llama 4 Maverick. Each the Llama 4 Scout and Llama 4 Maverics fashions at the moment are out there for obtain on the Meta web site and Hugging Face.
The brand new fashions draw inspiration from Chinese language AI startup DeepSeek, adopting a machine studying method generally known as Combination of Consultants. This system lets completely different components of the mannequin specialise in particular duties, resulting in improved efficiency and better effectivity.
As we speak is the beginning of a brand new period of natively multimodal AI innovation.
As we speak, we’re introducing the primary Llama 4 fashions: Llama 4 Scout and Llama 4 Maverick — our most superior fashions but and the perfect of their class for multimodality.
Llama 4 Scout
• 17B-active-parameter mannequin… pic.twitter.com/Z8P3h0MA1P
— AI at Meta (@AIatMeta) April 5, 2025
The Llama 4 Scout is the lighter, extra compact mannequin, highly effective sufficient to run on a single Nvidia H100 GPU. It’s designed for effectivity with out sacrificing efficiency. In distinction, Llama 4 Maverick is the heavy hitter—extra similar to superior fashions like GPT-4o and Gemini 2.0 Flash.
Meta has skilled its Llama 4 fashions to be natively multimodal by pre-training them on huge quantities of unlabeled textual content, picture, and video knowledge. In easy phrases, this implies the fashions can perceive and reply to each photos and textual content seamlessly. Therefore, they’re constructed to deal with conversations that contain extra than simply phrases.
Alternatively, the Meta remains to be coaching Llama 4 Behemoth, which CEO Mark Zuckerberg proudly calls “the very best performing base mannequin on this planet.” Internally, Meta refers to it as one of many smartest massive language fashions (LLMs) ever created.
Based on an organization, the Llama 4 Maverick mannequin is “the workhorse” of the 2 and excels at picture and textual content understanding for “common assistant and chat use instances”.
Llama 4 Scout comes with 17 billion lively parameters, supported by 16 consultants, bringing the entire parameter rely to 109 billion. Llama 4 Maverick additionally options 17 billion lively parameters however faucets right into a a lot bigger pool of 128 consultants, giving it a lift in efficiency and flexibility.
Nonetheless, the multimodal options of Meta AI are at present restricted to English customers within the US, that means no Ghibli fashion picture era with Meta AI simply but.
