Meta has launched Llama 4, a new set of AI models designed to enhance its AI assistant across platforms like WhatsApp, Messenger, and Instagram. This collection includes two main models: Llama 4 Scout and Llama 4 Maverick. Llama 4 Scout is compact enough to operate on a single Nvidia H100 GPU and boasts a context window of 10 million tokens, allowing it to effectively manage more information.
Meta claims that this model outperforms Google’s Gemma 3 and Gemini 2.0 Flash-Lite models across multiple benchmarks while maintaining its efficiency. Similarly, Llama 4 Maverick, positioned as a competitor to OpenAI’s GPT-4o and Google’s Gemini 2.0 Flash, reportedly achieves results comparable to DeepSeek-V3 in coding and reasoning tasks, all while using fewer active parameters.
Meta is also developing the Llama 4 Behemoth model, which the company claims will be the highest-performing base model available, featuring 288 billion active parameters and a total of 2 trillion parameters. Although this model has not yet been released, Meta anticipates that it will surpass its competitors, such as GPT-4.5 and Claude Sonnet 3.7, on key STEM benchmarks.
The Llama 4 series employs a “mixture of experts” (MoE) architecture, optimizing resource usage by activating only the necessary parts of the model for specific tasks. Meta’s upcoming LlamaCon conference on April 29 will further outline the company’s direction for AI models and corresponding products.
Despite labeling the Llama 4 collection as “open-source,” Meta faces criticism over its licensing terms, which require commercial entities with over 700 million monthly users to obtain permission for use. This stipulation has led some to argue that Llama is no longer truly open source, as noted by the Open Source Initiative in 2023.
