Meta Unveils Llama 4: A New Collection of AI Models
Meta has announced Llama 4, a new collection of AI models that power the Meta AI assistant on the web and in WhatsApp, Messenger, and Instagram. The new models are designed to improve the performance and capabilities of Meta’s AI assistant, and are available for download from Meta or Hugging Face.
Llama 4 Models
The Llama 4 collection includes two new models: Llama 4 Scout and Llama 4 Maverick. Llama 4 Scout is a small model that can fit in a single Nvidia H100 GPU, while Llama 4 Maverick is a larger model that is more akin to GPT-4o and Gemini 2.0 Flash. Meta is also working on a third model, Llama 4 Behemoth, which is expected to be the highest-performing base model in the world.
Performance Benchmarks
According to Meta, Llama 4 Scout has a 10-million-token context window and outperforms other models, including Google’s Gemma 3 and Gemini 2.0 Flash-Lite, as well as the open-source Mistral 3.1, across a broad range of widely reported benchmarks. Llama 4 Maverick also outperforms other models, including OpenAI’s GPT-4o and Google’s Gemini 2.0 Flash, and is comparable to DeepSeek-V3 in coding and reasoning tasks using less than half the active parameters.
Llama 4 Behemoth
Llama 4 Behemoth has 288 billion active parameters with 2 trillion parameters in total. While it hasn’t been released yet, Meta says Behemoth can outperform its competitors on several STEM benchmarks. Meta CEO Mark Zuckerberg has described Behemoth as the highest-performing base model in the world.
Architecture and Licensing
Meta has switched to a “mixture of experts” (MoE) architecture for Llama 4, which conserves resources by using only the parts of a model that are needed for a given task. The company plans to discuss future plans for AI models and products at its LlamaCon conference, which is taking place on April 29th. As with its past models, Meta calls the Llama 4 collection “open-source,” although the license restrictions have been criticized by some in the open-source community.
Conclusion
Meta’s Llama 4 collection is designed to improve the performance and capabilities of its AI assistant, and offers a range of new models and features. While the license restrictions have been criticized, the new models offer significant improvements in performance and capabilities.
FAQs
Q: What are the features of Llama 4 Scout and Llama 4 Maverick?
A: Llama 4 Scout is a small model that can fit in a single Nvidia H100 GPU, while Llama 4 Maverick is a larger model that is more akin to GPT-4o and Gemini 2.0 Flash.
Q: How does Llama 4 Scout compare to other models?
A: According to Meta, Llama 4 Scout outperforms other models, including Google’s Gemma 3 and Gemini 2.0 Flash-Lite, as well as the open-source Mistral 3.1, across a broad range of widely reported benchmarks.
Q: What is the purpose of Llama 4 Behemoth?
A: Llama 4 Behemoth is expected to be the highest-performing base model in the world, and will be used to power Meta’s AI assistant on the web and in WhatsApp, Messenger, and Instagram.
Q: What is the license restriction for Llama 4?
A: The Llama 4 license requires commercial entities with more than 700 million monthly active users to request permission from Meta before using its models.
