Meta releases a new collection of AI models, Llama 4, in the Llama family – on a Saturday, less.

There are four new models in general: Llama 4 scouts, llama 4 Maverick, and Llama 4 behemoth. Everyone is trained in “large amounts of unsold text, image, and video data” give them “broad understanding of visual,” said Meta.

Success of open models from Chinese Ai Treferkek, making par or better than the previous flagship of the meta llama, reportedly moving the development of llam. Meta said there were scrambled war rooms to determine how dereseeeek lowered the cost of running and deploying models like r1 and v3.

Scout and Maverick are expressly available at Llama.com and from Meta partners, including AI Dev Platch Hugging facing, while Behemoth is still on. Meta says Meta Ai, Ai-Powered Assistant in apps including WhatsApp, Messenger, and Instagram, updated to use Llama 4 in 40 countries. Multimodal features are limited to US in English for today.

Some developers can issue LLAMA 4 license.

Users and companies “inhabited” or in a “main area of ​​business” in the EU are prohibited by using or distributing AI and Data Privacy Privacy Requirements. .

“These models of Llama 4 mark the start of a new era for the llama ecosystem,” META wrote on a blog post. “This is the beginning for listing Llama 4.”

Meta says llama 4 is the first models of models to use a mixture of experts (moe) architecture, which is more effective in training and answering questions. Moe architectures are basically damaged data processing tasks in subtasks and then deliver them to small, specialists “experts” models.

Maverick, for example, there are 400 billion total parameters, but only 17 billion active parameters of total experts in 128 “.” (Parameters nearly matching the problem solving skills in a model.) Scout has 17 billion active parameters, 16 experts, and 109 billion overall parameters.

According to the internal test of Meta, Maverick, the company says best for “General Assistant and Chat” ‘s Geminies, Chat “‘ s Geminies, Chat’s Geminies, Gemini’s Gemini in Google 2.5 prie, anthropic’s Claude 3.7 Sonnet-4.5.

Scout strengths are in tasks such as the document summarizing and reasoning in many cenerebases. Unusual, it has a very large context window: 10 million tokens. (“Tokens” represents pieces of raw text – eg, the word “fantastic” divided by “mole in English,” tal-mution words

Scout can run into a NVIVIA H100 GPU, while Maverick requires a NVIDIA H100 DGX system, according to the meta.

The Unexpected Behemoth of the Meta will need even beefier hardware. According to the company, Behemoth has 288 billion active parameters, 16 experts, and almost two trillion total parameters. The internal meta benchmarking is the Behemoth who is gotten-4.5, Claude 3.7 Sonnet, and Gemini) in many stem measurement tests such as solving mathematical problems.

In the note, none of the Llama 4 models are a “Reason” model of Openia’s O1 and O3-Mini lines. Modeling rational models check their answers and usually respond to questions that are more reliable, but as a result of longer than traditional, “not justification” models to provide answers.

Interestingly, Meta says all the models of llama it will refuse to respond to “argue” questions less frequently. According to the company, Llama 4 answers the “debate” political and social topics of the previous harvest of Llama models. In addition, the company says, LLAMA 4 is “more balanced” that prompts this flat-out is not fun.

“(Y) IU will be able to trust (lllama 4) to provide helpful, true responses without judgment,” a meta tech talker told. “(W. Continue to make Llama more responsive to answer it with many questions, can answer different views (…) and not favoring some views of others.”

Those tweaks come as home alliances accusing AI political work.

Many close confidants of President Donald Trump, including Elon Musk and Crypto and AI “Czar” David Stans, say many chatbots Censor conservative perspective. History sacks sung The Chatgpt of Openia is specifically “programmed to wake up” and is unconscious about political sensitive topics.

Indeed, AI bias is an unstable technical problem. AI’s own AI, Xai, have struggling To create a chatbot without pushing some political views of others.

That does not prevent companies including Openia from Renewal Their AI models to answer many questions than before, in particular questions about controversial political subjects.