Meta is making significant strides in the AI space with the launch of its new Llama 4 models: Scout and Maverick. According to early tests, both models are outperforming competitors widely.
Scout is tailored for handling large documents, intricate requests, and extensive codebases. In contrast, Maverick excels in managing both text and visuals, making it ideal for smart assistants and chat interfaces.
These new models are available on Llama.com and through partnerships, including Hugging Face. Meta is integrating them into its AI assistant across popular apps like WhatsApp, Messenger, and Instagram in 40 countries, though initially limited to the U.S. and English-speaking users.
Under the hood, Llama 4 features a Mixture of Experts (MoE) setup, enhancing both the model’s training efficiency and response speed. Scout operates with 17 billion active parameters distributed across 16 expert modules, allowing it to outperform Google’s products like Gemma 3 and Gemini 2.0 Flash-Lite, while efficiently running on a single Nvidia H100 GPU.
One of Scout’s remarkable abilities is its capacity to process up to 10 million tokens, managing large volumes of text and visual data adeptly. Maverick, with 17 billion parameters across 128 expert networks, is also impressive, delivering competitive performance to OpenAI’s GPT-4 and Google’s Gemini 2.0 Flash.
However, it falls slightly short compared to Google’s Gemini 2.5 Pro and Anthropic’s Claude 3.7 Sonnet regarding top-tier results. Meta has hinted at developing an even more powerful model, Llama 4 Behemoth, expected to have 288 billion active parameters and potentially surpassing GPT-4.5 and Claude 3.7 in STEM-related tasks.
As Meta continues to enhance its AI capabilities, users can anticipate sharper responses, improved image generation, and more relevant advertising across its platforms.