Meta has launched the first two models from the new Llama 4 series — Scout and Maverick — designed for various applications in the field of artificial intelligence. These are open-source multimodal models, intended to handle complex tasks involving language and image processing, and they demonstrate advanced performance compared to competing models.
Llama 4 Scout is a relatively compact model, with 17 billion active parameters and 16 experts. It was designed for tasks such as summarizing multiple documents, analyzing large-scale user activity, and drawing conclusions from code. According to Meta, Scout is particularly fast, supports a context of up to 10 million tokens, and can run on a single GPU card.
Llama 4 Maverick, considered the main “workhorse” model among the two, also includes 17 billion active parameters but with 128 experts, and is intended for more general use in AI assistants and conversations. It excels in areas such as code, inference, long-context processing, multilingual support, and image understanding. According to Meta, it has a clear advantage over models like GPT-4o and Gemini 2.0 and successfully competes with DeepSeek v3.1 in inference and programming tasks.
Both models were developed as part of the process of building the Llama 4 Behemoth supermodel, which is still in training. According to Meta, Behemoth has 288 billion active parameters and is expected to be one of the most powerful foundation models in the world. Another model — Llama 4 Reasoning — is also expected to be revealed soon, focusing on advanced inference.
The new models are available for download via Meta’s website, the Hugging Face platform, and the cloud platforms of Azure, AWS, and Google Cloud, under a commercial and research license.