Meta takes out two Llama 4 AI models

Meta announced Llama 4, its new collection of AI models which now feed the Meta AI assistant on the web and in Whatsapp, Messenger and Instagram. The two new models, also available for download from Meta or Hugging Face, are Llama 4 Scout – a small model capable of “Integrating into a single NVIDIA H100 GPU” – and LLAMA 4 MAVERICK, which is more like GPT -4O and Gemini 2.0. Meta says he is still forming the Llama 4 giant, which, according to the Meta CEO, Mark Zuckerberg, is “the most efficient basic model in the world”.

According to Meta, Llama 4 Scout has a context window of 10 million billion – the working memory of an AI model – and beats Gemma 3 and Gemini 2.0 from Google Gemini, as well as the Mistral 3.1 of Open -Source 3.1, “on a wide range of largely reported benchmarks”, while installation in a single NVIDIA H100 GPU “. ». Meta makes similar claims on the performance of her larger Maverick model compared to the GPT-4O of Openai and Gemini 2.0 of Google, and says that its results are comparable to Deepseek-V3 in coding and reasoning tasks using “less than half of the active parameters”.

Meanwhile, the Llama 4 giant has 288 billion active parameters with 2 billions of parameters in total. Although it has not yet been published, Meta says that Behemoth can surpass its competitors (in this case GPT-4.5 and Claude Sonnet 3.7) “On several STEM references”.

For Llama 4, Meta says that it went to a “mixture of experts” (MOE), an approach that keeps resources using only the parts of a necessary model for a given task. The company plans to discuss future plans for AI models and products at its Llamacon conference, which takes place on April 29.

As with its previous models, Meta calls the Llama 4 collection “Open-Source”, although Llama was criticized for its license restrictions. For example, the Llama 4 license requires commercial entities with more than 700 million monthly active users to request Meta’s permission before using its models, which the Open Source initiative wrote in 2023 it withs the “open source” category.