Meta is released Llama 4, a new AI of AI models

Meta has released a new AI models collectionLlama 4, in the Llama family – Saturday, no less.

There are four new models in total: Llama 4 Scout, Llama 4 Maverick and Llama 4 Behemoth. Everyone was trained in “large amounts of non -marked text, image and video data” to give them a “broad visual understanding”, Meta says.

The success of the open models from the Chinese AI Lab Deepseek workshop, which perform the Par or Better from previous Meta Llama models, reportedly kicked the development of the blade in Overdrive. Meta is said to have mixed war rooms to decipher how Deepseek has reduced the cost of operating and developing models such as R1 and V3.

Scout and Maverick are openly available Llama.com And by Meta partners, including the face of the AI ​​Dev platform, while Behemoth is still in training. Meta says Meta Ai, AI’s assistant fueled by applications such as WhatsApp, Messenger and Instagram, have been informed to use Llama 4 in 40 countries. Multimodal characteristics are limited to the US in English for now.

Some developers may face Llama 4 permit.

Users and companies “inhabit” or with a “main place of business” in the EU is The use or distribution of models is prohibitedProbably the result of the governance requirements imposed by the laws on the privacy of the region and the data. (In the past, Meta has decapitated these laws as overly burdens.) In addition, as with the previous Lama releases, companies with more than 700 million monthly active users must seek special permission from META, which may grant or refuse to discredit it.

“These Llama 4 models mark the beginning of a new era for Llama ecosystem”, Meta wrote in a blog post. “This is just the beginning for Llama Collection 4.”

Meta says that Llama 4 is its first group of models to use a mixture of MOE architecture, which is more computational to training and answering queries. MOE architectures basically collapse data processing work to subjects and then transfer them to smaller, specialized “expert” models.

Maverick, for example, has 400 billion total parameters, but only 17 billion active Parameters in 128 “experts”. (The parameters correspond to approximately the problem -solving skills of a model.) Scout has 17 billion active parameters, 16 experts and 109 billion total parameters.

According to Meta’s internal tests, Maverick, whom the company says is better for the use of “general assistant and conversation”, such as creative writing, exceeds models such as Openai’s GPT-4O and Google’s Gemini 2.0 in a specific coding, reasoning, multilingual, long-lasting, but Maveric Gemini 2.5 Pro, Google’s Claude 3.7 Sonnet and Openai’s GPT-4.5.

The strengths of Scout are in tasks such as the summary and reasoning of the document in large codes. Unique, it has a very large window of frame: 10 million chips. (“The chips” represent pieces of raw text – eg the word “fantastic” divided into “fan”, “tas” and “tic.”) In simple English, scout can take images and up to millions of words, allowing it to process and work with extremely long documents.

Scout can run in a single GPU Nvidia H100, while Maverick requires a NVIDIA H100 DGX or equivalent system, according to Meta calculations.

Meta’s non -circulated Behemoth will even need Beefier material. According to the company, Behemoth has 288 billion active parameters, 16 experts and nearly two trillion total parameters. META’s internal comparative assessment has exceeded the Behemoth GPT-4.5, Claude 3.7 Sonnet and Gemini 2.0 Pro (but not 2.5 Pro) in various ratings that measure STEM skills such as mathematical problem solving.

It is worth noting that none of the Llama 4 models are a proper “reasoning” model according to the O1 and O3-MINI lines of Openai. Reasoning models control their answers and generally respond to more reliable questions, but therefore need more time than traditional, “non-dimensional” models to provide answers.

Interestingly, Meta says it coordinates all Llama 4 models to refuse to answer less often to the “controversial” questions. According to the company, Llama 4 responds to “discussed” political and social issues that the previous cultivation of Lama models would not be. In addition, the company says that Llama 4 is “dramatically more balanced” with which it urges flat-out will not have fun.

“[Y]ou can calculate [Lllama 4] To provide useful, real replies without crisis, “a Meta spokesman told TechCrunch.”[W]We continue to make the blade more sensitive to answer more questions, can respond to a variety of different views […] And it does not favor some opinions for others. ”

These bites come as some White House allies accuse AI Chatbots of being very politically “awakened”.

Many of President Donald Trump’s close confidentialists, including billionaire Elon Musk and Crypto and AI “Czar” David Sacks, have claimed that the popular conservative views of AI Chatbots censor. Sacks have historical separated Openai’s chatgpt as “scheduled to wake up” and false about the political issue.

In fact, bias in AI is an intolerable technical problem. Musk’s own AI company Xai struggled to create a chatbot that does not support any political views on others.

This has not stopped companies, including OpenAI to adapt AI models to answer more questions than they would have previously, specific questions about controversial issues.