SAN FRANCISCO — Meta has announced the release of the first models in the Llama 4 series. The company said the release markes a significant advancement in AI technology.
These models are designed to enable the creation of more personalized multimodal experiences. The flagship models, Llama 4 Scout and Llama 4 Maverick, are now available for download on llama.com and Hugging Face. These models are the first open-weight, natively multimodal models with unprecedented context length support. They are also the first to utilize a mixture-of-experts (MoE) architecture.
Llama 4 Scout: This 17 billion active parameter model, equipped with 16 experts, is the most powerful multimodal model in its class. It outperforms all previous Llama models and surpasses competitors like Gemma 3, Gemini 2.0 Flash-Lite, and Mistral 3.1 on a wide range of benchmarks. Despite its power, Llama 4 Scout fits on a single NVIDIA H100 GPU.
Llama 4 Maverick: With 128 experts, this 17 billion active parameter model is the best in its class. It surpasses GPT-4o and Gemini 2.0 Flash on multiple benchmarks. Llama 4 Maverick achieves comparable results to the new DeepSeek v3 on reasoning and coding tasks, but with less than half the active parameters. It offers an industry-leading context window of 10 million and an experimental chat version that scored an ELO of 1417 on LMArena.
Llama 4 Behemoth: The Teacher Model
Meta also previewed Llama 4 Behemoth, a 288 billion active parameter model with 16 experts. This model is currently in training and is expected to be one of the smartest large language models (LLMs) in the world. It has already outperformed GPT-4.5, Claude Sonnet 3.7, and Gemini 2.0 Pro on several STEM benchmarks, including MATH-500 and GPQA Diamond.
The Llama 4 models are designed for both performance and efficiency. The MoE architecture allows for more efficient training and inference, delivering higher quality results within a fixed training FLOPs budget compared to traditional dense models.
Meta emphasizes the importance of openness in driving innovation. By making Llama 4 Scout and Llama 4 Maverick available to the public, Meta aims to empower developers, enterprises, and AI enthusiasts to build new experiences and integrate these models into various workflows.
Meta believes that the most intelligent systems should be capable of taking generalized actions, conversing naturally with humans, and solving complex problems. The company is committed to further research and development, with plans to share more about its vision at LlamaCon on April 29.
Users can try Meta AI built with Llama 4 in WhatsApp, Messenger, Instagram Direct, and on the Meta.AI website starting today. The models will also be available through Meta’s partners in the coming days.
Meta’s Llama 4 models represent a new era for AI personalization. They offer multimodal intelligence at a compelling price point and outperform models of significantly larger sizes. As Meta continues to innovate, the potential for these models to enhance products and services across various industries is immense.
