LLAMA
Meta has just released a new set of open-source AI models under the Llama 4 banner, Scout, Maverick, and Behemoth.
Trained on massive amounts of text, image, and video data, they’re designed to better understand and respond across different formats.
The release seems to have been fast-tracked after Chinese AI lab DeepSeek launched high-performing models that matched or outpaced earlier Llamas.
Reports suggest Meta scrambled internal teams to figure out how DeepSeek made their models more efficient and cheaper to run - Zuck was spooked, ya’ll.
Scout and Maverick are now publicly available on Meta’s site and through platforms like Hugging Face. Behemoth, the largest of the bunch, is still in training.
Meanwhile, Meta’s own AI assistant, used in WhatsApp, Messenger, and Instagram, has been updated with Llama 4 in 40 countries.
That said, visual features are currently limited to English and only available in the US.
What sets these models apart is their use of “mixture of experts” (MoE) architecture.
GPU or go home
Rather than using the full model each time, only selected parts, called experts, are activated, saving computing power.
Maverick has 400 billion total parameters, but only 17 billion are used per task.
Scout is smaller, with 109 billion total and the same number of active parameters.
Scout runs on a single Nvidia H100 GPU, making it more accessible for smaller developers.
Maverick requires more advanced hardware, while Behemoth, still being fine-tuned, needs heavy-duty systems.
Meta says Behemoth is showing strong performance on technical tasks like maths and coding, outperforming several leading models.
There are licensing limits to be aware of.
Anyone based in the EU or with a main business address there is blocked from using or sharing the models, likely due to regional data laws.
A few things to know:
Llama 4 uses MoE tech to boost efficiency without maxing out compute resources.
Licences are restricted for EU users and companies with massive user bases.
Scout and Maverick are live now, while Behemoth is still under development.
Companies with over 700 million users also need special approval from Meta.
Meta’s also made changes to how the models respond to tricky topics.
Llama 4 is designed to be less hesitant around political or social questions and more consistent in how it handles different viewpoints.
This update follows claims, mostly from US political circles, that chatbots are biased or censor certain views.
Meta says its aim is to give more helpful answers across a wider range of topics, without taking sides.
It's worth noting that none of these are “reasoning” models like OpenAI’s o1 series, which are slower but better at double-checking their answers.
Side note, Scout and Maverick sound like AI cowboys, and honestly, I’m here for it.