Meta launches Llama 4 - A suite of new competitive multimodal AI models

Llama Scout and Maverick are now available, while Llama Behemoth remains pending.

The battle for AI dominance is intensifying, and the past few months have seen a flurry of activity. After significant updates from Gemini, ChatGPT, and Deepseek, Meta has entered the fray with a compelling suite of AI models. The social media giant's parent company has launched Llama 4, positioning it as a strong contender in the competitive AI landscape.

Earlier today, Meta announced the rollout of an updated AI strategy through its blog, Meta AI, featuring the new Llama 4 models. These models will be integrated into the Meta AI assistant across the web and within popular social media platforms like WhatsApp, Instagram, and Messenger, aiming to enhance user experiences.

Meta introduced two key models: Llama Scout, a remarkably efficient model designed to operate within a single NVIDIA H100 GPU. The company touts it as the "best multimodal model in its class," delivering superior results compared to Gemma 3, Gemini 2.0 Flash-Lite, and Mistral 3.1 across a wide range of benchmarks.

The second model, Llama Maverick, boasts 17 billion active parameters and a total of 400 billion parameters. It has demonstrated exceptional performance, surpassing Deepseek v3 in reasoning and coding tasks. Both Llama Scout and Llama Maverick are accessible through Llama Meta and Hugging Face.

Image: Meta

Looking ahead, Meta is previewing Llama Behemoth, which CEO Mark Zuckerberg describes as "the highest performing base model in the world." This model, with its impressive 288 billion active parameters and 2 trillion total parameters, is projected to outperform Claude Sonnet 3.7, Gemini 2.0 Pro, and GPT 4.5 on instruction-tuning benchmarks.

While Meta has presented Llama 4 as an 'open-source' offering, the licensing terms warrant careful consideration. The Llama 4 license incorporates specific limitations, particularly for large-scale commercial applications. Businesses with over 700 million monthly active users are required to obtain prior authorization from Meta before deploying these models.

A pivotal element of Meta's Llama 4 release is its new architectural design: A 'mixture of experts' (MoE) framework. This method allows the AI to conserve computational resources by selectively engaging only the necessary parts of the model for each specific task, resulting in greater efficiency. This strategic architectural change is set to be a focal point at Meta's upcoming LlamaCon event on April 29th, where the company is anticipated to reveal its future roadmap for AI models and products.

What are your thoughts on this? Please share your expressions in the comments section below