Meta Unveils 3 New Llama 4 AI Models

Meta has announced the launch of its new Llama 4 models. The lineup includes Llama 4 Scout, Llama 4 Maverick, and a preview of Llama 4 Behemoth, each made to give developers tools for creating personalized multimodal experiences. These models are now available for download on llama.com and Hugging Face, and they can be used in Meta platforms like WhatsApp, Messenger, Instagram Direct, and the web.

Llama 4 Scout has 17 billion active parameters and 16 experts, which allows it to run on a single Nvidia H100 GPU. It can handle a large context of up to 10 million tokens, making it useful for tasks like summarizing several documents and analyzing big codebases.

Llama 4 Maverick also has 17 billion active parameters but includes 128 experts, which makes it powerful for reasoning, coding, and multilingual tasks. Meta says both models outperform competitors such as GPT-4o and Gemini 2.0 Flash-Lite across various benchmarks.

The previewed Llama 4 Behemoth is called Meta's most powerful model yet, with 288 billion active parameters and nearly two trillion total parameters. It excels in STEM-focused benchmarks like MATH-500 and GPQA Diamond and serves as a teacher model for distilling smaller versions.

Meta says the models use a mixture of expert architectures to enhance computational efficiency. They are trained with FP8 precision and new reinforcement learning strategies to improve reasoning and coding abilities. The company has also added safeguards to reduce response bias and ensure application safety.