Meta has announced a major leap forward in the development of its open-source language models with the release of Llama 4. This latest update to the LLM (Large Language Model) family includes two new models: Llama 4 Maverick and Llama 4 Scout, both made available under a specific open-source licence with defined usage conditions. With this release, Meta is placing a strong focus on multimodality and efficiency, while also challenging the relevance of the RAG (Retrieval-Augmented Generation) approach that has become widely adopted in the AI community.
A Significant Shift in Architecture
Rather than offering a mere iteration of Llama 3.3, Meta has chosen to overhaul the architecture entirely. Previously based on the conventional transformer model, Llama 4 now adopts a Mixture of Experts (MoE) approach — a strategy that has gained traction over the past year, notably through Mixtral 8x22B by Mistral AI and DeepSeek R1. MoE models work by using different “experts” — specialised sub-networks within the larger model — with only a small portion of the model’s weights activated for each input token.
This selective activation significantly boosts computational efficiency, allowing the model to maintain high performance while drastically reducing the energy required for inference. The shift towards MoE represents a broader trend in AI development aimed at creating scalable and environmentally conscious models that don’t sacrifice capability.
Two Powerhouse Models with Different Roles
The Llama 4 Scout model is equipped with 109 billion parameters, of which only 17 billion are active during inference. This makes it a relatively lightweight and energy-efficient choice for real-time applications or environments with limited resources. In contrast, Llama 4 Maverick is a more heavyweight contender, with 400 billion total parameters, though it shares the same 17 billion active parameters at inference time. This setup provides a balance between model depth and efficiency, enabling advanced tasks while keeping compute costs under control.
Both models are designed with enterprise-level deployment in mind, offering high levels of performance and scalability. Their release under open-source licences underscores Meta’s continued commitment to pushing forward AI innovation in a transparent and collaborative way.
Multimodality and the RAG Debate
Another key element of the Llama 4 release is the emphasis on multimodal capabilities — the ability of models to process and understand information across different formats, such as text, images, and potentially even audio. This reflects a growing shift in AI development towards building systems that can mimic the richness of human understanding by integrating diverse forms of input.
In addition, Meta is challenging the dominance of RAG as the primary technique for enhancing language models with external knowledge. RAG has been widely used to improve accuracy by combining model outputs with information pulled from external databases or documents. However, Meta argues that its new architecture and training methods may reduce the need for such augmentation, paving the way for more autonomous and context-aware models.
A Strategic Release with More to Come
The unveiling of Llama 4 Scout and Maverick is likely just the beginning. Meta has indicated that more models in the Llama 4 line-up are in development and could be released soon. Each model is expected to cater to specific use cases, giving developers and researchers greater flexibility in choosing the right tool for their needs.
With Llama 4, Meta is not just competing in the global AI race — it’s reshaping it. By delivering open-source models that combine architectural innovation, energy efficiency, and multimodal capability, the company is positioning itself at the forefront of next-generation artificial intelligence.
This bold step highlights Meta’s broader strategy: to democratise access to cutting-edge AI tools, support open collaboration, and accelerate the pace of innovation across the industry. As more details emerge and adoption grows, Llama 4 could well become a defining force in the evolution of global AI.