Meta has announced Llama 4, its latest collection of AI designs that currently power the Meta AI aide on the web and in WhatsApp, Carrier, and Instagram. Both brand-new designs, additionally offered to download and install from Meta or Hugging Face, are Llama 4 Precursor– a little version with the ability of “suitable in a solitary Nvidia H100 GPU”– and Llama 4 Radical, which is much more comparable to GPT-4o and Gemini 2.0 Flash. Meta states it’s still in the procedure of training Llama 4 Leviathan, which Meta Chief Executive Officer Mark Zuckerberg says is “the highest possible carrying out base version on the planet.”
According to Meta, Llama 4 Precursor has a 10-million-token context home window– the functioning memory of an AI version– and defeats Google’s Gemma 3 and Gemini 2.0 Flash-Lite designs, along with the open-source Mistral 3.1, “throughout a wide series of commonly reported standards,” while still “suitable in a solitary Nvidia H100 GPU.” Meta makes comparable cases concerning its bigger Radical version’s efficiency versus OpenAI’s GPT-4o and Google’s Gemini 2.0 Flash, and states its outcomes approach DeepSeek-V3 in coding and thinking jobs making use of “much less than half the energetic specifications.”
At The Same Time, Llama 4 Leviathan has 288 billion energetic specifications with 2 trillion specifications in overall. While it hasn’t been launched yet, Meta states Leviathan can exceed its rivals (in this situation GPT-4.5 and Claude Sonnet 3.7) “on numerous STEM standards.”
For Llama 4, Meta states it switched over to a “blend of professionals” (MoE) design, a technique that saves sources by utilizing just the components of a version that are required for an offered job. The business intends to go over future prepare for AI designs and items at its LlamaCon seminar, which is taking place on April 29th.
Just like its previous designs, Meta calls the Llama 4 collection “open-source,” although Llama has actually been slammed for its certificate limitations. As an example, the Llama 4 license calls for business entities with greater than 700 million month-to-month energetic individuals to demand authorization from Meta prior to utilizing its designs, which the Open Resource Effort wrote in 2023 takes it “out of the group of ‘Open up Resource.'”
发布者：Wes Davis，转转请注明出处：https://robotalks.cn/meta-releases-two-llama-4-ai-models/
