Meta launches world's most powerful AI model Llama 4
Llama 4: The Scout model has 17 billion active parameters and is equipped with 16 experts. It is capable of running on a single Nvidia H100 GPU. At the same time, the Maverick model also comes with 17 billion parameters but has 128 experts, which makes its depth and analysis capability much more powerful.
Meta has unveiled the initial preview of its new Llama 4 series of AI models to the world. The company has made two cutting-edge multimodal capabilities featuring open-source models Llama 4 Scout and Llama 4 Maverick available to the open community.
Meta claims that these models are the first open AI models based on the Mixture-of-Experts (MoE) architecture. Compared to earlier versions, features like better context window, excellent power efficiency, and faster processing have been added to these models. Along with this, the company also gave a glimpse of its largest AI creation Llama 4 Behemoth, which is considered to be the most powerful model to date.
According to Meta's official blog post, both these models are completely open-source and can be downloaded from Hugging Face or Meta's official Llama website. Users can now experience these models on WhatsApp, Messenger, Instagram Direct, and the Meta.AI website.
The Scout model has 17 billion active parameters and is equipped with 16 experts. It is capable of running on a single Nvidia H100 GPU. At the same time, the Maverick model also comes with 17 billion parameters but has 128 experts, making its depth and analysis capability much more powerful.
According to Meta's internal tests, the Maverick model outperformed high-end models such as Gemini 2.0 Flash, DeepSeek v3.1, and GPT-4o on several parameters including MMMU (image reasoning), ChartQA (chart understanding ability), GPQA Diamond (knowledge and reasoning), MTOB (ability to handle long contexts). At the same time, the Scout model also outperformed models such as Gemma 3, Mistral 3.1, and Gemini 2.0 in its category.
