Meta launches Llama 4 series, its most advanced AI models yet

Meta has officially launched its most advanced AI models to date with the Llama 4 series.

The Llama 4 series includes Llama 4 Scout and Llama 4 Maverick, both of which feature 17 billion active parameters and a mixture-of-experts architecture that boosts efficiency when it comes to understanding text and images.

Scout is designed to run on a single H100 GPU and offers a whopping 10 million token context window. That’ll come in handy for tasks like summarising long documents and dealing with complex data sets.

Maverick, meanwhile, is a larger model with 128 experts (and a total of 400 billion parameters). It outperforms the likes of OpenAI’s GPT-4o and Google’s Gemini 2.0 across a range of benchmarks, including reasoning and coding tasks.

Both of these models are distilled from the larger Llama 4 Behemoth, which is still in training. Behemoth has 288 billion active parameters and is said to outperform GPT-4.5 and other leading models in tests focused on STEM subjects.

You can download the new models from llama.com and Hugging Face, while they’re also integrated into Meta’s AI assistant that spans WhatsApp, Instagram and more.

Meta is committed to open-source AI and has been leading the charge among the big tech companies. However, there are concerns that licensing restrictions for users in the EU could undermine those efforts due to the continent’s stringent data privacy laws.

The launch comes as Meta seeks to establish itself as a leader in the generative AI space, particularly after its rivals have enjoyed so much success. The company plans to host LlamaCon on April 29 to discuss its AI strategy further.

With the evolution of multimodal AI systems continuing apace, Meta is also working on reducing bias in the responses given by its AI systems.

The Llama 4 series is a significant step forward.