Meta Unveils Llama 4, an AI Model with a 10 Million Token Context Window
Department: Review
Meta has taken a major step forward in AI with the release of the next generation of its language models, Llama 4. The Llama 4 Scout model stands out in this series, being able to process context on up to **10 million tokens, making it one of the most advanced AI models on the market.
AMP Version
What's new in Llama 4?
Llama 4 includes three models:
- Llama 4 Scout
- Llama 4 Maverick
- Llama 4 Behemoth (still in development)
Llama 4 Scout: incredible amount of context
The flagship of the line — Llama 4 Scout — is able to analyze huge amounts of information thanks to a context window for 10 million tokens. This is especially useful when processing long documents, videos, and large sets of images (up to 48 per request).
Thanks to the Mixture of Experts (MoE) architecture, the model activates only the necessary blocks of parameters, which allows for efficient use of computing resources. Llama 4 Scout performs well on benchmarks, beating competitors including Google Gemma 3 and Mistral 3.1.
“We introduce Llama 4 Scout and Llama 4 Maverick, the first natively multimodal open-weight models with unprecedented context length support, as well as our first models built using the Mixed Experts (MoE) architecture. We also introduce Llama 4 Behemoth, one of the smartest LLMs in the world and the most powerful yet to serve as a teacher for our new models,” writes Meta
Llama 4 Maverick and Behemoth: Performance and Scale
The Maverick model shows comparable performance to GPT-4o and DeepSeek-V3 on encoding and reasoning tasks, while using fewer active parameters.
While Llama 4 Behemoth hasn't been released yet, it's already on everyone's lips: it will include 288 billion parameters and is designed to handle the most complex tasks. According to Meta, Behemoth outperforms GPT-4.5 and Claude 3.7 on STEM benchmarks.
Product Integration and License
Llama 4 AI models have already been implemented in Meta's popular apps — WhatsApp, Messenger, and Instagram, opening up new possibilities for interacting with AI in real time.
However, it is worth noting that despite being open source, Llama 4's license limits commercial use to companies with over 700 million users, which has caused controversy in the open-source community.
Llama 4 Scout and Llama 4 Maverick are available for download on llama.com and Hugging Face. The AI models are also available on WhatsApp, Messenger, Instagram DM, and the Meta.AI website.
You can also check out our detailed review about language models
What's next?
Details about the development of Llama 4 and Meta's future plans are promised to be revealed at the LlamaCon event, which will take place on April 29.
Comments