Meta has unveiled three new models from its artificial intelligence (AI) software Llama 4, as the Facebook and Instagram parent company aims to have a stake in the generative AI (GenAI) race.
Two of the models are now available but not its most powerful Llama 4 model, which Meta teased would be released at a later date. CEO Mark Zuckerberg said that the so-called Llama 4 Behemoth would be "the highest performing base model in the world".
While that model is still in training, here is what we know about the other two AI models that are now available for download.
Llama 4 Scout
Of the three Llama 4 models announced, Scout is the smallest but is designed to be the speediest.
Scout has 109 billion total parameters, which are seen to be the building blocks of the AI that value and determine how the model processes information and generates outputs. The more they have, the more complex tasks they can handle.
It also has a Mixture of Experts (MoE), which is like a team of specialists and each expert has a specific task, such as maths. Scout has 16 experts.
It also has a 10 million token context window, which means that it can digest about 8 million English words and then summarise them.
Meta says that Scout can handle multi-document summarisation, parsing extensive user activity for personalised tasks, and reasoning over vast codebases.
Llama 4 Maverick
The larger Maverick Llama 4 model boasts 128 experts and 400 billion total parameters.
Meta said in a blog post that it is the best-in-class multimodal model, meaning it can simultaneously process different modalities - such as text, video, audio, and image - to generate outputs.
Meta also said that it exceeded comparable models like GPT-4o and Gemini 2.0 on coding, reasoning, multilingual, long-context, and image benchmarks, and it's competitive with the much larger DeepSeek v3.1 on coding and reasoning.
Llama 4 Behemoth
Though it is not out yet, Meta said that Behemoth has 16 experts and nearly two trillion total parameters.
The company said it offers state-of-the-art performance for non-reasoning models on maths, multilinguality, and image benchmarks.
But according to the publication Venture Beat, when using different reasoning models to assess Behemoth against DeepSeek’s R1 and OpenAI’s o1 models, Meta’s offering was not always the clear winner in some cases.
Open weight
In his announcement video on Instagram, Zuckerberg said that Meta’s "goal is to build the world’s leading AI, open source it, and make it universally accessible so that everyone in the world benefits… I’ve said for a while that I think open source AI is going to become the leading models, and with Llama 4, that is starting to happen".
However, the company has called the Llama 4 models open weight and not open source, after receiving criticism that it does not match the Open Source Initiative’s official definition.
This is mainly because Meta does not reveal the data that the AI models were trained on.
Is Llama 4 available in Europe?
It is not available to customers based in Europe, which includes not being available for research or personal use.
Meta has not specified why; however, it halted the launch of previous Llama models in the EU due to regulatory concerns, as the data used to train its models comes from Meta users.
But if you are living outside of Europe, you can download the models on Meta or Hugging Face.
Changing political favour
Meta says that all Llama 4 models show substantial improvement on "political bias".
The company said that "specifically, [leading large luggage models] historically have leaned left when it comes to debated political and social topics," and that Llama 4 is more inclusive to right-wing politics, which follows Zuckerberg’s embracing of US President Donald Trump after his 2024 election win.
