Meta has introduced Llama 4, a new series of advanced AI models within its Llama family. The collection includes four models: Llama 4 Scout, Llama 4 Maverick, and Llama 4 Behemoth. These models were trained using extensive unlabeled text, image, and video data to enhance their visual understanding capabilities. The development of Llama 4 was reportedly accelerated due to the success of DeepSeek’s open models, which performed on par or better than Meta’s previous flagship models. Scout and Maverick are available to the public through Llama.com and Meta’s partners, while Behemoth remains in training. Meta AI, the company’s assistant across platforms like WhatsApp and Messenger, now utilizes Llama 4 in 40 countries, though multimodal features are limited to the U.S. in English. The Llama 4 license restricts usage by entities domiciled in the EU and requires special permission for companies with over 700 million monthly active users. Meta highlights that Llama 4 is its first set of models to use a mixture of experts (MoE) architecture, improving computational efficiency. Maverick, with 400 billion total parameters, excels in general assistant tasks, surpassing models like OpenAI’s GPT-4o and Google’s Gemini 2.0 in certain benchmarks. Scout, featuring a large context window of 10 million tokens, is strong in document summarization and reasoning over large codebases. Behemoth, still unreleased, boasts 288 billion active parameters and nearly two trillion total parameters, outperforming several models in STEM-related evaluations. Meta has also adjusted Llama 4 to respond more frequently to contentious questions and provide balanced, factual responses.
— new from TechCrunch