The finish has Lama 4 was announcedHis latest collection of AI models, which now feed the meta and assistant on the Internet And in WhatsApp, Messenger and Instagram. Two modern models, also available for download with finish Or HuggingThis is Llama 4 Scout-Mała model capable of “matching in a single GPU NVIDIA H100” -I Llama 4 Maverick, which is more similar to GPT-4O and Gemini 2.0 Flash. Meta claims that there is still 4 Behemoth in Lama training, whose meta CEO Mark Zuckerberg says This is “the best -acting basic model in the world”.
According to Meta, Llama 4 Scout has a 10 million context window-working model AI-I overcomes Google Gemma 3 and Gemini 2.0 Flash-Lite, as well as “Mistral Open Source 3.1”, “in a wide range of widely reported tests”, and still “in one NVIDIA H100 GPU respectively.” The meta claims similar claims about the performance of a larger Maverick model compared to the GPT-4O OpenNAI and Google’s Gemini 2.0 Flash, and claims that its results are comparable to Deepseek-V3 in coding and reasoning using “less than half of the active parameters”.
Meanwhile, Lima 4 Beemoth has 288 billion busy parameters from a total of 2 trillion parameters. Although it has not been released yet, Meta claims that Behemoth can surpass its competitors (in this case GPT-4.5 and Claude Sonnet 3.7) “on several STEM tests”.
In the case of Lama 4, META claims that it has changed to the architecture of the “expert mix” (MOE), an approach that retains resources using only part of the model needed for a given task. The company plans to discuss future plans of AI models and products at the Llamacon conference, which will take place on April 29.
As in the case of previous models, the meta called the Lama 4 “Open Source” collection, although Llama has been criticized for licensing restrictions. For example Call license 4 Requires commercial entities with over 700 million busy users to ask for permission from the finish he wrote in 2023 It takes this “from the” Open Source “category.