“10 million tokens and new architecture. Meta has released the Flagman models of Shi LLAMA 4 April 7, 12:35 to share: Llama 4 already integrated in Meta AI (photo: Meta) Author: Anastasia Meta Meta Zuckerberg presented the first models of new LLAMA 4, some 3.7 Sonnet from anthropic. There are three”, – WRITE ON: ua.news
Llama 4 is already integrated into Meta AI (photo: META)
Meta Mark Zuckerberg presented the first models of the new Llama family 4, some of which exceed the best competitors’ products, such as GPT-4.5 from Openai and Claude 3.7 Sonnet from Anthropic.
Three new models have been offered in total: Llama 4 Scout, Llama 4 Maverick and Llama 4 Behemth. Scout and Maverick are openly available on llama.com and Meta partners, including AI Hugging Face Developer Platform, whereas Behemth is still taught.
Meta says Llama 4 is the first series of models that use Mixture of Experts architecture (Moe). MOE architecture basically break the tasks of processing data on subtasks and then delegate them to smaller specialized “Expert “to models.
New trend. Openai and Meta copy deekeeek approach in attempts to create cheap and effective shi
Maverick, for example, has 400 billion parameters, but only 17 billion active parameters in 128 Experts. Scout has 17 billion active parameters, 16 “experts” and 109 billion total parameters.
According to Meta’s internal testing, Maverick is best suited for use as a common assistant. The model can help with tasks such as creative writing. Meta argues that this AI exceeds models such as GPT-4O from Openai and Gemini 2.0 from Google, in certain tests. However, Maverick is lagging behind more powerful recent models, such as Gemini 2.5 Pro from Google, Claude 3.7 Sonnet by Anthopic and GPT-4.5 from Openai.
Quality issues. Zuckerberg compared his si with much cheaper models from Deepeseek
The strengths of Scout are tasks such as generalization of documents and reasoning on large code bases. AI has a very large context window: 10 million tokens. The model can also take images up to millions of words, which allows you to work with extremely large documents.
In this case, according to META calculations, Scout can work on one NVIDIA H100 graphic processor. MAVERICK requires NVIDIA H100 DGX or equivalent. An unspecified Meta Behemth model will require even more powerful equipment. According to the company, Behemoth has 288 billion active parameters, 16 “experts” and almost two trillions of general parameters. Internal comparative analysis Meta shows that Behemth is superior to GPT-4.5, Claude 3.7 Sonnet and Gemini 2.0 Pro (But not 2.5 Pro) in several tests, in particular to solve mathematical problems.
Meta says Meta AI, its artificial intelligence assistant in applications, including WhatsApp, Messenger and Instagram, has been updated for the use of LLAMA 4 in 40 countries. But multimodal features are currently available for US users and only English. In Ukraine, this chat bot is currently unavailable.
Altman is not delighted. Meta AI chat can get a separate program
Earlier it was reported that the success of the open models from the Chinese DEEPSEEK artificial intelligence laboratory, which operate at or better than the previous flagship models of Meta Llama, forced Mark Zuckerberg to speed up developments. It was reported that Meta has created separate teams to decipher how Deeceek reduced the cost of launching and deploying models such as R1 and V3.