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

Have meta AI has published a new collection of modelsLama 4, in its Lama family – Saturday, no less.

There are four new models in total: Lama 4 Scouts, Lama 4 Mavarick and Lama 4 Behemoth. Meta says that everyone was trained in “large -scale labeled text, images and video data” to give “wide visual understanding”.

Success of Models Open from Chinese AI Lab DipscIt is reported that Llama Vikas has been kicked in overdrive, which is equal or better than the previous flagship Lama models of Meta. It is said that DPCIC has reduced the cost of running and deploying models is thought to have meta scrambled War rooms to understand R 1 And V3The

Scout and Mavarick are publicly available Llama.com And the AI ​​Dev platform hugs from Meta partners with the face, while Behemoth is still in training. Meta says its AI-powered assistant Meta AI AI has been updated to use Lama 4 in applications including WhatsApp, Messenger and Instagram. Multimodal features are now limited to English.

Some developer Lamas may issue issues with 4 licenses.

User and agencies, including “residence” or in the EU, including a “original place of business” Models are prohibited to use or distributeProbably the result of the administration’s requirements imposed by the AI ​​and Data Privacy Act of this region. (In the past, there is meta These laws have reduced In addition, as a burden.) In addition, like the previous Lama Release, more than a monthly active users of more than 1 million monthly active users must be requested for a special license from the meta, which Meta can be granted or denying on its only discretion.

“These Lama 4 models identify the launch of a new era for the Lama Ecosystem,” Meta Wrote in a blog postThe “It’s just starting for the collection of Lama 4″ ”

Meta says that the Lama 4 is a combination of the first models for using a mixture (MOE) architecture, which is more proficient for training and answering questions. MOE architecture basically breaks data processing tasks in subtasks and then represents small, special “specialist” models.

For example, Mavarick has 400 billion total parameters but only 17 billion Active Parameters across 128 “experts.” (Parameters are fairly adjusting to a model problem solving skills)) The scout has 17 billion active parameters, 16 experts and 109 billion total parameters.

According to Meta’s internal tests, Maverick, which the company says, is used to use cases as creative writing for “general assistant and chat”, outdoo GPT -4O And Google Gemini 2.0 Specific coding, logic, multilingual, long-line and image criteria. However, Mavarick does not perfectly measure recent models like Google Gemini 2.5 ProEthnographic CLOD 3.7 SonnetAnd Openai’s GPT -4.5The

The power of the scout is in the shortage of the document and the tasks like logic on the larger codebase. Uniquely, it has a very large context window: 10 million token. (Tokens present “bits of raw texts – such as” fantastic “the word” fan, “” TAS “and” tick “) plain can take scout images and up to several million words, allowing it to process and work with extremely long documents.

Scouts can run on a single nvidia H 100 GPU, while an NVEDIA H1 100 DGX system or equivalent is needed, according to Mater Count.

Meta’s unpublished Behemoth even needs Biffier hardware. According to the agency, Behmoth has 20 billion active parameters, 1 16 experts and about two trillion total parameters. As a result of Meta’s internal benchmarking, there are discounts on various evaluation of stem skills such as solving mathematics problems.

Significantly, none of the Lama 4 models is a correct “argument” model in OpenAe’s line O 1 And O3-miniThe Reasonable models are answering their answers and usually reliably react to questions, but as a result it takes longer than the traditional, “non-resoning” model to provide answers.

The funny thing is, Meta says it often tunes the 4 model to refuse to answer the “controversial” question. According to the agency, Lama 4 responds to “controversial” to political and social issues that the previous crop of the Lama models is not. Furthermore, the agency says Lama 4 is “dramatically more balanced” with which it does not entertain flat-out.

“[Y]You can count [Lllama 4] Helpful without verdict, to provide practical reactions, “Meta spokesperson told TechCrunch.”[W]Continues to make Lama more responsive so that it answers more questions, react to different aspects of different aspects […] And not in favor of some opinion on others. “

These tweets come as some of the White House allies are accused of “waking up” very politically as AI Chattabot.

Billionaire Elon Mask and Crypto and AI “Caesar” David Sacks Many President Donald Trump’s close believer complained that popular AI Chatbots Sensor conservative aspectThe The sack is the histor tihassically Out Openai’s Chatzept “has been programmed to wake up” and is unfaithful about political issues.

In fact, the bias in AI is a complex technical problem. Kasturi has its own AI company Jai Fight To create a chatbot that does not support some political views on others.

It did not turn off companies including Openai Consistency Their AI models are to answer more questions in special questions related to controversial issues.