Meta has announced the latest additions to its AI family with the launch of Llama 4 models, designed to power its chatbot across platforms like WhatsApp, Instagram, and Messenger. The new models are being positioned as strong contenders against OpenAI’s ChatGPT and Google’s Gemini.
Meet Llama 4: Scout, Maverick and Behemoth
Meta has introduced two new downloadable models—Llama 4 Scout and Llama 4 Maverick—now available on the Meta website and Hugging Face. The company also gave a sneak peek into its most advanced model yet: Llama 4 Behemoth, which it claims is among “the smartest LLMs in the world.”
While Behemoth is currently in preview as a “teacher model”, Scout and Maverick are available for use and are expected to play a major role in Meta’s consumer AI tools.
Multimodal by Design: Text + Image understanding
A standout feature of the Llama 4 models is their native multimodal capability. Unlike earlier versions, these new models are trained on text, images, and videos, allowing them to seamlessly interpret and respond across formats—although for now, multimodal features are limited to English users in the US.
What’s new in Llama 4 Scout and Maverick?
Llama 4 Maverick
- 17 billion active parameters
- 128 experts
- Described as a “product workhorse”, ideal for general chat, creative writing, and precise image understanding.
Llama 4 Scout
- 109 billion total parameters
- 17 billion active parameters
- 16 experts
Optimized for document summarization, code reasoning, and long-context tasks, featuring a 10 million token context window.
Meta says Scout outperforms top models like Gemma 3, Gemini 2.0 Flash Lite, and Mistral 3.1 on various benchmarks.
Inspired by DeepSeek: Using “Mixture of Experts”
Meta has employed a machine learning technique called Mixture of Experts (MoE), inspired by Chinese AI startup DeepSeek. This allows different parts of the model to specialize in specific tasks, resulting in more efficient and powerful performance without increasing latency.
What about reasoning capabilities?
Unlike OpenAI's o3-mini or DeepSeek R1, the current Llama 4 models are not yet reasoning models. These models don’t attempt human-like logical thinking but instead focus on speed, multimodality, and creative output.
Where can you use Llama 4?
You can now access the Llama 4-powered Meta AI on:
- Messenger
- Meta AI’s website (in over 40 countries)
However, advanced multimodal features—like generating Studio Ghibli-style images—are currently exclusive to US-based English users.