n Saturday, Meta released Llama 4 as well as their new collection of AI models. Llama 4 Scout, Llama 4 Maverick, and Llama 4 Behemoth are the
new models that have been launched and are trained on “large amounts of unlabeled text, image, and video data”. In order to offer them a “broad visual understanding,” as Meta reported.
As previous Meta performance has been overpowered by the Chinese AI model DeepSeek, Meta’s product Llama development has been driven into overdrive. Even more so, Meta is said to have scrambled war rooms in order to decipher the process of how DeepSeek lowered the cost of running as well as deploying models such as R1 and V3.
The new models presented by Meta, meaning Scout and Maverick, are widely available on Llama.com and from Meta’s partners. The AI development platform Hugging Face is now available, yet Behemoth is still under training.
Meta has also reported that MetaAI and its AI-powered assistant, which is available across apps such as WhatsApp, Messenger, and Instagram, have been updated to use Llama 4 in 40 countries. However, multimodal features are limited only to the U.S. in English for now.
Even more so, in a blog post, Meta wrote, “These Llama 4 models mark the beginning of a new era for the Llama ecosystem,” adding, “This is just the beginning for the Llama 4 collection. Users and companies located or with a “principal place of business” located in the UE are for now prohibited from using or even distributing the models.
Subscribe to our newsletter
Meta also reported that the Llama 4 is its first cohort model that uses a mix of experts architecture that is more computationally efficient compared to training and answering queries. MoE (mix of experts) architectures basically break down data processing into subtasks and then delegate them to even smaller specialized “expert” models.
Maverick has over 400 billion total patterns, only using 17 billion active parameters across 128 “experts”. Even more so, Scout has 17 billion active parameters that are used on 16 experts and 109 billion total parameters.
More so, Scout can run on a single Nvidia H100 GPU, while Maverick needs an Nvidia H100 DGX system or equivalent, as Met calculated.
A spokesperson reported for TechCrunch “[Y]ou can count on [Lllama 4] to provide helpful, factual responses without judgment,” also adding “[W]e’re continuing to make Llama more responsive so that it answers more questions, can respond to a variety of different viewpoints […] and doesn’t favor some views over others.”.
It is also worth mentioning that many of Donald Trump’s close confidants, including Elon Musk and David Sacks, the AI and crypto “czar,” have alleged that popular AI chatbots censor conservative views.
.webp)
