Meta has released a new collection of him, Llama 4, in her family Llama – on Saturday, no less.
There are four new models in total: Llama 4 Scout, Llama 4 Maverick, and Llama 4 Behemoth. They were all trained in “large amounts of text data, image and unmarked videos” to give them “broad visual understanding,” Meta says.
The success of the models open from the Chinese Lab Deepseek, which perform at the same time or better than the previous models of the Llama Llame flag, is reported to begin the development of the Llam in overdrive. Meta is said to have introduced war rooms to decipher how Deepseek reduced the cost of running and setting models such as R1 and V3.
Scout and Maverick have been opened at llama.com and by meta partners, including the AI Dev Hugging Face platform, while Behemoth is still in training. Meta says Meta, her assistant with the energy in app, including WhatsApp, Messenger and Instagram, has been updated to use Llama 4 in 40 countries. Multimodal characteristics are limited to the US in English for now.
Some developers can deal with Llama 4 license.
Users and companies “set” or with a “main business site” in the EU are forbidden to use or distribute models, likely the result of governance requirements established by the region and data intimacy laws. (In the past, Meta has decreed these laws as too severe.) In addition, as with previous lament omissions, companies with more than 700 million monthly active users must require a special license from Meta, which Meta can give or deny at its sole discretion.
“These Llama 4 models mark the beginning of a new era for the Llama ecosystem,” Meta wrote in a blog post. “This is just the beginning for the Llama 4.” collection ”
Meta says Llama 4 is its first set of models to use a mixture of expert architecture (MOE), which is more effective in a computing mode for the questions and answers. MOE architectures essentially break down the database tasks under the underlying and then delegate 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 of solving the problem of a model.) The scout has 17 billion active parameters, 16 experts and 109 billion total parameters.
According to Meta’s internal testing, Maverick, which the company says is best for “general assistant and chat” cases of use of creative writing, exceeds models such as Openai’s GPT-4o and Google’s 2.0 for some codification, reasoning, multi-language, long taxes and images. However, Maverick is not measured with more capable models such as Gemini 2.5 pro the Google, Claude 3.7 of Anthropic, and GPT-4.5 of OpenAi.
The Stout strengths stand tasks such as the document compilation and the reasoning on the large basics of the codes. Uniquely, it has a very large context window: 10 million signs. (“Tokens” represent pieces of raw text – eg the word “fantastic” divided into “fan”, “tas” and “tic”) in plain English, the scout can take images and up to millions of words, allowing it to process and work with extremely long documents.
Scout can operate on a single NVIDIA H100 GPU, while Maverick requires a Nvidia H100 DGX or equivalent system, according to meta calculations.
Mahemoth inaccessible Meta will need even better hardware. According to the company, Behemoth has 288 billion active parameters, 16 experts and nearly two trillion total parameters. Meta internal behemotion has behemoth exceeding GPT-4.5, Claude 3.7 sonnet, and Gemini 2.0 pro (but not 2.5 pro) in some ratings that measure stem skills such as solving mathematics problems.
Attention, none of the Llama 4 models is a proper “reasoning” model along the O1 O3 and O3-MINI lines. Reasoning patterns control their answers and generally answer more reliable questions, but consequently last longer than traditional, non-reasoning models to answer.
Interestingly, Meta says she allocated all her models Llama 4 to refuse to answer “controversial” questions less frequently. According to the company, Llama 4 responds to the “debated” political and social topics that the previous Llama models would not be. In addition, the company says, Llama 4 is “dramatically more balanced” with which it makes it flat will not entertain.
“(Y) You can rely on (llala 4) to provide useful, actual non -judgmental answers,” said a Meta spokesman in Techcrunch. “(W) is continuing to make the varnish more responsible in order to answer more questions, can answer a variety of different views (…) and does not favor some views on others.”
These changes come after some White House allies accuse him of chatbots of being too “woke up” politically “.
Many of the close believers of President Donald Trump, including billionaire Elon Musk and Crypto and the “Czar” David Sacks, have claimed that popular chatbots censor conservative chatbots. Sacks has historically singled out Openai’s chatgt as “programmed to wake up” and not true about the political topic.
In fact, prejudice in it is an interactive technical problem. Music’s own company, Xai, has struggled to create a chatbot that does not adopt some political views on others.
This has not stopped companies including Openai from fixing their models to answer more questions than they would have before, in particular questions about controversial subjects.
