Meta has unveiled its latest AI model collection, Llama 4, featuring four versions: Llama 4 Scout, Llama 4 Maverick, and Llama 4 Behemoth. These models were trained using a vast array of unlabeled text, images, and videos for enhanced visual comprehension. The launch was reportedly accelerated by competition from DeepSeek, a Chinese AI lab with models exceeding previous Llama versions.
Scout and Maverick are publicly available, while Behemoth is still in development. The updated Meta AI assistant now incorporates Llama 4 across various apps, although features for non-English speakers are currently limited. The Llama 4 license restricts EU-based users and requires large companies to seek special permissions from Meta.
Llama 4 introduces a more efficient mixture of experts (MoE) architecture, allowing specialized tasks to be processed by dedicated models. Maverick boasts 400 billion total parameters but operates with 17 billion active parameters across 128 experts. Scout excels at document summarization with a 10 million token context window and can run on a single Nvidia H100 GPU.
Behemoth, with 288 billion active parameters, is designed for advanced STEM tasks. While Llama 4 models are not classified as reasoning models, they have been fine-tuned to address contentious topics more freely. Meta aims for these updates to enhance response quality and balance in answering varied viewpoints, addressing criticism of AI bias amid political debates.
