Meta releases Llama 4, a new crop of flagship AI models

Meta released a new group of AI, Llama 4 models, in the Llama family – on Saturday, no less.

There are four new models in Total: Llama 4 Scout, Llama 4 MAVERICK and Llama 4 Beheemoth. All of them were trained on “large quantities of texts, image and video data that are not named” to give them “a wide visual understanding,” says Meta.

It is said that the success of the open models of AI LAB Deepseek, which leads to equality or better than the previous pioneering Llama models in Meta, and it was reported that it kicked the development of Llama to Overdrive. It is said that Meta has a whipped war rooms to decode how to lower Deepseek from the operating cost and publish models such as R1 and V3.

Scout and MAVERICK are openly available on Llama.com and Meta partners, including the AI ​​Dev Facting Face platform, while Beheemoth is still in training. Meta says that Meta AI, the assistant with which Amnesty International works via applications including WhatsApp, Messenger and Instagram, has been updated to use Llama 4 in 40 countries. The multimedia features are limited to the United States in English at the present time.

Some developers may face a problem with Llama 4 license.

Users and companies are “custom” or “major workplace” in the European Union are probably prohibited to use or distribute models, as a result of the requirements of governance imposed by data privacy laws in the region. (In the past, Meta denounced these laws as very stressful.)

Meta wrote in a blog post: “These models from Llama 4 represent the beginning of a new era of the ecological system,” Meta wrote in a blog post. “This is just the beginning of the Llama group 4.”

Meta says that Llama 4 is the first set of models to use a combination of expert structure (MEE), which is more effective in terms of mathematical training and response to inquiries. MEE structures mainly dismantle data processing tasks to sub -tasks and then delegate them to smaller and specialized “experts” models.

Mavrick, for example, has 400 billion teachers, but only 17 billion active Teachers through 128 “experts”. (Parameters are almost compatible with the problem solving skills in the form.) Scout has 17 billion active teachers, 16 experts and 109 billion parameters.

According to Meta’s internal test, MAVERICK, which the company says is the best for “general assistant and chat” cases such as creative writing, exceeds models such as GPT-4O from Openai and Gemini 2.0 from Google on some coding, thinking, traditional multiple standards and long comfort. However, Maverick does not fully measure modern models more capable like Google’s Gemini 2.5 Pro, Claude 3.7 Sonnet, and Openai’s GPT-4.5.

Scout’s strengths lie in tasks such as summarizing documents and thinking about large code rules. Uniquely, it has a very large context window: 10 million icons. (The symbols represent parts of the raw text – for example, the word “wonderful” is divided into “fan”, “tas” and “TIC”) in the normal English, the scout can take pictures and even millions of words, allowing it to process and work with very long documents.

Scout can work on the NVIDIA H100 graphics processing unit, while MAVERICK requires the NVIDIA H100 DGX system or its equivalent, according to Meta accounts.

The Meta giant, which has not even released, will need beamier devices. According to the company, bethmoth owns 288 billion active teachers, 16 experts and about trillion from the total parameters. The interior standards of Meta have a giant that outperforms GPT-4.5, Claude 3.7 Sonnet and Gemini 2.0 Pro (but not 2.5 Pro) on many assessments that measure STEM skills such as solving mathematics problem.

It should be noted that none of the Llama 4 models is the appropriate “thinking” model similar to Openai’s O1 and O3-MINI. Dismantling the models, verify their answers and generally respond to questions more reliable, but as a result, it takes longer than the “non -modified” traditional models to provide answers.

Interestingly, Meta says she has seized all Llama 4 models to refuse to answer “controversial” questions often. According to the company, Llama 4 responds to political and social topics “the discussion” that the previous crop will not do from Llama. In addition, the company says, Llama 4 “is very much more balanced” and which is demanded by a flat that will not enjoy it.

“[Y]Ou can be relied upon [Lllama 4] “To provide useful realistic responses without a ruling,” Meta spokesman told Techcrunch.[W]He continues to find Lama more responsive to answer more questions, she can answer a variety of different views […] Some opinions are not preferred over others. “

These amendments come at a time when some of the White House allies accuse artificial intelligence of “waking up” politically.

Many close close associates of president Donald Trump, including billionaire Elon Musk, Karva and Ai “Caesar” David, claimed that the opinions of preserving Amnesty International. Sacks historically dedicated Chatgpt from Openai as “programmed to wake up” and is not true to the political issue.

In fact, prejudice to artificial intelligence is an intractable technical problem. MUSK has struggled with AI, Xai, to create a tension that does not support some political opinions on others.

This did not prevent companies, including Openai, from controlling their artificial intelligence models to answer more questions than they would have previously asked, especially questions related to controversial topics.