Mark Zuckerberg says Meta's latest Llama models put open-source AI in the driving seat
Meta Platforms Inc. unveiled the latest additions to its Llama family of large language artificial intelligence models on Saturday, claiming that they're among the most powerful ever released to the public.
Advanced Llama 4 Models
The new models, part of the Llama 4 series, are now accessible through the Meta AI assistant on various platforms such as web, Messenger, WhatsApp, and Instagram. They can also be downloaded from Meta directly or via Hugging Face.
The Llama 4 series includes Llama 4 Scout, a small model designed to fit inside a single Nvidia H100 graphics processing unit, and Llama 4 Maverick, a larger model comparable to OpenAI’s GPT-4o and Google LLC’s Gemini 2.0 Flash, both with 17 billion active parameters.
Future Release: Llama 4 Behemoth
Meta's Chief Executive, Mark Zuckerberg, revealed that they are still developing the largest model in the Llama 4 lineup, named Llama 4 Behemoth. Once released, it is expected to be the "highest performing base model in the world."
Advancements in Modality
Meta claims that the Llama 4 models are the most advanced and "best in class" in terms of modality, capable of processing various data formats such as text, images, audio, and video. This enables them to comprehend complex scenarios and generate improved responses.
AI Agent Capabilities
Chris Cox, Meta's Chief Product Officer, mentioned that the Llama 4 models are designed to power AI agents with enhanced reasoning skills capable of completing tasks autonomously and with minimal supervision.
Innovative Technology: Mixture-of-Experts
Meta utilized a new system called "Mixture-of-Experts" (MoE) to create the Llama 4 models, enhancing efficiency by activating only the necessary parts for specific tasks. This system enables faster processing and reduced energy consumption.
Focus on Safety and Balance
Aside from performance, Meta emphasized making the Llama 4 models safer and more balanced to prevent biased or harmful responses. This ensures that the models provide well-rounded answers, especially in controversial or politically sensitive topics.
Upcoming Developments
Looking ahead, Meta is set to host its first annual LlamaCon AI conference on April 29, where they may unveil the Llama 4 Behemoth model and introduce a standalone Meta AI application.
Overall, Meta's latest Llama models represent a significant advancement in open-source AI technology, bringing cutting-edge capabilities to the forefront of AI development.
