Meta presented the first two models from the Llama 4 multimodal family: Llama 4 Scout and Llama 4 Maverick. There will be four models in the family in total. Two of them have already been launched, and the rest will appear later. According to Meta, all of them were trained on a large amount of unannotated text, images, and videos to gain extensive visual understanding capabilities.
The reason for the accelerated launch of new models was the successful promotion of open models of the Chinese laboratory DeepSeekthat perform on par or better than previous Meta models. According to reports, Meta has urgently organized «war rooms» to figure out how DeepSeek could reduce the cost of launching and using models like R1 and V3.
According to the company, Llama 4 Maverick — is the «workhorse» that does the best job of understanding images and text for «typical assistant and chat tasks». Llama 4 Scout, on the other hand, is suitable for tasks such as summarizing multiple documents, analyzing a large number of user actions for personalized tasks, and logically processing large code bases.
Meta also announced Llama 4 Behemoth — a future model that the company calls «one of the smartest AI models in the world». CEO Mark Zuckerberg added that the fourth model — Llama 4 Reasoning — will be introduced within the next month.
Llama 4 Maverick and Scout are already available for download on the Llama website and on Hugging Face, and are already integrated into Meta AI — including WhatsApp, Messenger, and Instagram (in messages).
Llama 4 Scout has 17 billion active parameters and 16 experts. According to Zuckerberg, this model is «extremely fast» and natively multimodal. It has an almost infinite context of up to 10 million tokens. At the same time, it is designed to run on a single GPU.
Llama 4 Maverick also has 17 billion active parameters, but already 128 experts. Meta claims that Maverick:
This model is still in the training phase, but Meta already claims that it will be «the most powerful base model in the world», with 288 billion active parameters. Although it hasn’t been presented yet, we should expect more information about it in the near future.
We can also expect to learn more about the Llama 4 Reasoning model. After all, a major Meta conference for AI developers — LlamaCon will take place in a few weeks.
In January 2025, during the trialprocessit turned out that Mark Zuckerberg’s Meta forLlama AI traininghas illegally used millions of books.
Source: techcrunch, engadget