Meta is Released a new collection of AI modelsLama 4, in her Lama family – on Saturday, no less.
There are a total of four new models: Lama 4 Scouts, Lalma 4 Mauric, and Lalma 4 Behmot. Meta says everyone was trained in “a large quantity of labeled text, icon and video data” to give “a widely visual understanding”.
The success of the open model from the Chinese AI lab DepressicWhich performs better than the previous flagship lma models of the meta, allegedly kicked Lama Development in the over drive. It is said that Meta has turned around the war rooms to understand how DiPsic reduced the cost of running and deploying models like models R1 And v3.
Scouts and Morocco are openly available llama.com And from Meta’s partners, including the AI Dave platform, the face of embracing the platform, while Behmot is still in training. Meta says its AI -powered Assistant Meta AI in apps, including WhatsApp, Messenger, and Instagram, has been updated to use Lilama 4 in 40 countries. Multi -modal features are still limited to the United States in English.
Some developer Lama can take a problem with 4 licenses.
Consumers and companies in the European Union are with “Domicile” or “Principal Place of Business” Models are forbidden to use or distributePossibly the result of rule requirements imposed by the region’s AI and data privacy laws. (In the past, the meta is Dismissed these rules As is excessively burdensome.) In addition, like the previous lama release, companies with more than 700 million monthly active consumers will have to request a special license from meta, which Meta can give or refuse at its discretion.
Meta, “These Lalma 4 models indicate the beginning of a new era for the Lalma environmental system.” Written in a blog post. “This is the beginning of a collection of Lama 4.”
Meta says its first group model to use Lama 4 experts (MOE) architecture, which is more comparative for questions training and answering questions. MOE Architecture mainly breaks data processing works in all tasks and then entrusts them with small, special “specialist” models.
For example Morcock has 400 billion total parameters, but only 17 billion Rich Parameters in 128 “experts”. (The parameters are consistent with the skill of solving almost a model problem.) The scout has 17 billion active parameters, 16 specialists, and 109 billion total parameters.
According to Meta’s internal testing, Moric, which the company says is best for creative writing such as “General Assistant and Chat” issues, is higher than models like Openi. GPT-4O And Google’s Gemini 2.0 Some coding, reasoning, multi -linguistic, long context and image standards. However, the more unable to measure more recent models like Mauric Google Gemini 2.5 ProAnthropic Claude 3.7 SwantAnd Openi’s GPT-4.5.
The powers of the scout are in tasks such as the documentation summary and the reasoning on the larger codes base. Individually, it has a huge context window: 10 million tokens. –
According to the Meta Calculation, the scout can run on the same NVIDIA H100 GPU, while Maurque needs to be equal to the NVIDIA H100 DGX system or its equivalent.
Behmut of Meta will also need behemoot hardware. According to the company, Behemut has 288 billion active parameters, 16 experts, and about two trillion total parameters. Meta’s internal bench marking has been better performed on numerous diagnoses to measure stem skills such as GPT 4.5, Claude 3.7 Sonate, and Gemini 2.0 Pro (but not 2.5 Pro) solving mathematics problem.
Note, none of the Lama 4 models is an appropriate “reasoning” model on the openings O1 And O3-mini. The arguments models actually examine their answers and usually answer questions more reliable, but the resulting answers takes more time than the traditional, “irrational” model.
Interestingly, Meta says it reduced all its Lama 4 models to refuse to answer “controversial” questions. According to the company, Lilama 4 responded to the “debate” political and social topics that will not crop the back of the Lalma model. In addition, the company says, the Lama is 4 “dramatically more balanced”, which will not indicate it to be flat.
“[Y]You can trust [Lllama 4] A Meta spokesperson told the Tech Crunch, “Without a decision, a helper to provide a response to the facts.[W]Lama is continuing to make it more responsible so that it can answer more questions, respond to a variety of different perspectives. […] And does not support some ideas about others.
Those occasions come when some White House allies have accused Ai Chat Boats of “awakening” politically.
Many close acknowledgment of President Donald Trump, including billionaire Elon Musk and Crypto and Ai “Zar” David Sacks, have alleged that popular AI Chat Boats Censorship of Conservative thoughts. The sacks are historically Single out Open’s Chat GPT as a “awakening program” and dishonest about the political topic.
In fact, prejudice in AI is a complex technical problem. Musk’s own AI company, Zee Struggle To make a chat boot that does not confirm some political views about others.
It has not stopped companies including Openi To adjust In the special questions related to controversial articles, their AI model is to answer more questions, in special questions.