Meta unveils the cutting-edge Large Language Model (LLM) Llama 4, the latest iteration which Mark Zuckerberg, CEO of Meta, touts as the most powerful foundational AI model in the world today.
Llama 4 consists of two sub-models: Llama 4 Scout with parameter size of 17B, 16 sub-models (Experts) totaling 109B, described by Meta as the most efficient blended model currently available, capable of running on a single NVIDIA H100 GPU with 10M input support, outperforming similar models like Gemma 3, Gemini 2.0 Flash-Lite, and Mistral 3.1.
Another model, Llama 4 Maverick with a parameter size of 17B, 128 sub-models totaling 400B, is hailed as the best-performing blended model to date, boasting test results above GPT-4o and Gemini 2.0 Flash. Additionally, Meta highlights superior test results compared to DeepSeek v3 in reasoning and code-writing, with parameters operating at less than half the capacity.
Meta states that both models unveiled pull elements from a larger AI model than Llama 4 Behemoth, with a parameter size of 288B, 16 sub-models totaling 2T, delivering peak performance. Test results in STEM fields surpass GPT-4.5, Claude Sonnet 3.7, and Gemini 2.0 Pro, yet this model remains in the training phase and has not been released.
Llama 4 Scout and Llama 4 Maverick are open-source models available for download on llama.com and Hugging Face. For general users, Llama 4 updates are rolling out through Meta AI across WhatsApp, Messenger, Instagram Direct, and websites.
TLDR: Meta introduces the state-of-the-art Llama 4 LLM, comprising Scout and Maverick models with impressive performance surpassing previous AI models. Updates are available for download and being implemented across Meta’s platforms.
Leave a Comment