Meta released the LLAMA 4 collection, an AI model, on Saturday.
There are four new models: LLAMA 4 Scout, Llama 4 Maverick and LLAMA 4 Behemoth. Meta says he has been trained for “large amounts of unspoken text, images and video data” to provide “extensive visual understanding.”
The success of an open model of the PAR or better performance of META’s previous flagship lamar model is that the success of the open model of the open model of the CHINESE AI LAB Deepseek has begun to develop the LLAMA. META is known to scrambled the war room to decipher how deeper the cost of operating and distributing models such as R1 and V3.
Scouts and Mavericks can be purchased openly on LLAMA.COM, and Behemoth is still training in meta -partners, including the AI Dev platform hug. Meta said that META AI, an AI -based assistant, including Whatsapp, Messenger and Instagram, has been updated to use LLAMA 4 in 40 countries. Multimodal is currently limited to American English.
Some developers may have problems with LLAMA 4 licenses.
The user and company of the EU are prohibited from using or distributing models with “domestic” or “major business locations”, and it is likely to be a result of governance requirements imposed under the AI and Data Privacy Act in this area. (In the past, the META is too burdensome.) As with the previous LLAMA release, more than 700 million monthly active users must request META for a special license, which can be granted or rejected by META alone.
Meta said in a blog post, “This Lama 4 model means the beginning of the new era of the Rama ecosystem.” This is just the beginning of the Lama 4 collection. ”
META is the first model cohort to use an expert (MOE) architecture, which is more computational in training and responses. The MOE architecture basically decomposes data processing tasks into sub -tasks and delegates as a smaller, specialized “expert” model.
For example, Maverick has a total of 400 billion parameters, but only 1.7 billion. Active 128 “experts” parameters. (Parameters correspond to the problem -solving technology of the model.) The scout has 17 billion active parameters, 16 experts and a total of 109 billion total parameters.
According to META’s internal test, Maverick is the best suitable for “general auxiliary and chat” use, such as Creative Writing, and Maverick is a model such as OPENAI’s GPT-4O and Gemini 2.0 of Gemini 2.0. Excess. However, Maverick does not measure the recent models more than the same as Google’s Gemini 2.5 Pro, Anthropic’s Claude 3.7 Sonnet and Openai’s GPT-4.5.
The strength of the scout is in the same tasks as inferences for document summaries and large codebases. Originally, it has a very large context window: 10 million tokens. (“Token” represents primitive text. For example, the word “fantastic” was divided into “fans”, “TA” and “TIC”.) In ordinary English, scouts can shoot and work in very long documents by shooting images and millions of words.
The scout can run on a single NVIDIA H100 GPU, while Maverick can run according to the NVIDIA H100 DGX system or equivalent calculation.
The huge giant, which is not announced, requires more powerful hardware. According to the company, Behemoth has 28.8 billion active parameters, 16 experts and almost 2 trillion parameters. META’s internal benchmarking is more performance than GPT-4.5, Claude 3.7 Sonnet and Gemini 2.0 Pro (not 2.5 Pro) in several evaluations that measure STEM technologies such as solving math problems.
It is noteworthy that either of the LLAMA 4 model is not an appropriate “reasoning” model along the O1 and O3-Mini lines of Openai. The reasoning model confirms the facts and usually responds more stably to the question, but as a result, it takes longer to convey the answer than the traditional “non -class” model.
Interestingly, Meta says that he refused to answer all the “controversial” questions by adjusting all the Rama 4 models. According to the company, LLAMA 4 responds to the “discussed” political and social themes that are not the crops of the previous LLAMA model. LLAMA 4 also says that you can’t make a judgment by being “dramatically balanced.”
The meta spokesman said in TechCrunch, “(Y) OU can rely on (llama 4).” (W) makes Rama more responding to more questions, responds to a variety of perspectives, and does not prefer views on different perspectives.
This adjustment comes when the White House allies accused the AI chatbot too politically “awakened.”
The close reliability of President Donald Trump, including the billionaire Elon Musk and Crypto and AI “Czar” David Sacks, insisted on a popular AI chatbot censorship conservative view. SACKS has historically selected Openai’s CHATGPT as “programming to wake up” and not true about political topics.
In fact, AI prejudice is a technical problem that is difficult to deal with. XAI, a MUSK’s AI company, struggled to create a chatbot that does not support other people’s political views.
As a result, a company, including Openai, adjusted the AI model, and did not answer more questions than previous questions about controversial topics.