Meta Llama 4, a new crop of pioneering artificial intelligence models

Meta has A new collection of artificial intelligence models releasedLama 4, in the Lama 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.

The success of the open models from the Chinese AI Laboratory DibsicThat leads to equal or better than the former pioneering Llama models in Meta, and it was reported that it kicks 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.

Scouts and Track are publicly available on Llama.com From Meta partners, including the AI ​​Dev Huging Face platform, while the giant 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 “resident” or with the “main workplace” in the European Union are It is prohibited from using or distributing modelsIt is more likely as a result of the requirements of governance imposed by data privacy laws in the region. (In the past, Meta has Criticize these laws As very exhausted.

“These Llama 4 models represent the beginning of a new era for the ecological system,” Meta He 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, goes beyond models such as Openai’s GPT-4O And Google’s Gemini 2.0 On some coding, thinking, multi -language standards, long context, and images standards. However, Maverick does not measure more capable modern models like Google’s Gemini 2.5 ProAntarbur Claude 3.7 SonataAnd 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) can rely on (LLLAMA 4) to provide useful realistic responses without judgment,” a Meta spokesman told Techcrunch. (W) continues to make Lama more responsive to answer more questions, she can answer a variety of different views (…) and do not prefer some views 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 and Karva and Ai “Caesar” David, claimed that the famous chat tools are artificial intelligence Monitor conservative views. The bags are historically It was customized Chatgpt from Openai as “programmed to wake up” and is not sincere in the political issue.

In fact, prejudice to artificial intelligence is an intractable technical problem. Musk AI, Xai, has Struggle To create Chatbot does not support some political opinions on others.

This did not stop companies, including Openai adjust Their artificial intelligence models to answer more questions than they would have previously presented, especially questions related to controversial topics.