Meta, the parent company of Facebook, has unveiled two new AI models under its LLaMA 4 series. These models will now function as Meta AI assistants across platforms like WhatsApp, Messenger, and Instagram.
What Are the New LLaMA 4 Models?
Meta has released two new models: LLaMA 4 Scout and LLaMA 4 Maverick. Both of these models can be downloaded from Meta or Hugging Face.
- LLaMA 4 Scout is a smaller model designed to fit on an Nvidia H100 GPU.
- LLaMA 4 Maverick is comparable to GPT-4O and Gemini 2.0 Flash in terms of performance.
Meta’s Ambitious AI Plans
Mark Zuckerberg, Meta’s CEO, has shared that the company is in the process of training LLaMA 4 Behemoth, claiming it to be “the highest-performing base model in the world.”
- LLaMA 4 Scout features a context window of 10 million tokens, which is a metric used to measure an AI model’s memory capacity. This model is more powerful than Google’s Gemini 3 and the open-source Mistral 3.1.
Performance of LLaMA 4 Maverick
The LLaMA 4 Maverick model is also touted to surpass OpenAI’s GPT-4O and Google’s Gemini 2.0 Flash in terms of performance. It is expected to be at the same level as models like DeepSick-V3 in coding and reasoning tasks.
The Scale of LLaMA 4 Behemoth
With 288 billion active parameters and a total of 2 trillion parameters, LLaMA 4 Behemoth is expected to outperform its competitors, such as GPT-4.5 and Cloud Sonet 3.7, in various benchmarks related to Science, Technology, Engineering, and Mathematics (STEM).
LLaMA 4’s “Mixture of Experts” Architecture
Meta has adopted a “Mixture of Experts” (MoE) architecture for LLaMA 4. This allows the model to use only the necessary parts for a given task, thereby conserving resources while improving performance.
Meta’s Future AI Plans
Meta plans to discuss further AI models and upcoming products during the LLaMACon conference on April 19. This conference will focus on the company’s future AI developments and innovations.
Let me know if you need more details!
