Meta Unveils Llama 4: Advanced AI Models Launched

Meta just dropped a bombshell on a Saturday, unveiling the latest addition to its Llama family: Llama 4. This new lineup includes Llama 4 Scout, Llama 4 Maverick, and Llama 4 Behemoth. These models have been trained on heaps of unlabeled text, images, and videos, giving them a pretty solid grasp on visual stuff, according to Meta.

The push to release Llama 4 came after the Chinese AI lab DeepSeek showed off their models, which were either matching or beating Meta's previous Llama models. This apparently sent Meta into a frenzy, setting up war rooms to figure out how DeepSeek managed to make their models like R1 and V3 cheaper to run and deploy.

You can get your hands on Scout and Maverick right away from Llama.com or through Meta's partners like Hugging Face. But, Behemoth is still cooking in the lab. Meta's AI assistant, which you can find in apps like WhatsApp, Messenger, and Instagram, has been upgraded with Llama 4 in 40 countries. However, the fancy multimodal features are only available in the U.S. and in English for now.

There's a catch, though. If you're in the EU or your business is based there, you're out of luck because you can't use or share these models. This is probably because of the strict AI and data privacy laws over there. Meta's not a fan of these rules, by the way. Also, if your company has over 700 million monthly active users, you'll need to beg Meta for a special license, which they might just say no to.

"These Llama 4 models are just the start of something big for the Llama ecosystem," Meta boasted in their blog post. "We're just getting started with Llama 4."

Llama 4 is the first to use a mixture of experts (MoE) architecture, which makes training and answering questions a lot more efficient. It's like breaking down a big job into smaller tasks and handing them off to specialized "expert" models.

Take Maverick, for example. It's got a whopping 400 billion parameters in total, but only 17 billion are active across 128 experts. Scout, on the other hand, has 17 billion active parameters, 16 experts, and 109 billion total parameters.

Meta's internal tests show that Maverick, which is great for general assistant and chat tasks like creative writing, beats models like OpenAI's GPT-4o and Google's Gemini 2.0 in certain areas like coding, reasoning, and handling multiple languages and long texts. But, it doesn't quite stack up against newer, more powerful models like Google's Gemini 2.5 Pro, Anthropic's Claude 3.7 Sonnet, and OpenAI's GPT-4.5.

Scout shines in tasks like summarizing documents and working with large codebases. It's got a huge context window of 10 million tokens, which means it can handle images and millions of words at once. Scout can run on a single Nvidia H100 GPU, while Maverick needs a beefier Nvidia H100 DGX system or something similar.

As for Behemoth, it's still under wraps but will need some serious hardware. It's got 288 billion active parameters, 16 experts, and nearly two trillion total parameters. Meta's tests show it outperforming GPT-4.5, Claude 3.7 Sonnet, and Gemini 2.0 Pro in STEM skills like math problem solving, but it doesn't beat Gemini 2.5 Pro.

None of these Llama 4 models are what you'd call "reasoning" models like OpenAI's o1 and o3-mini. Those models double-check their answers and are more reliable, but they take longer to respond.

Interestingly, Meta says they've tweaked Llama 4 to be less likely to dodge "contentious" questions. The new models are more willing to tackle debated political and social topics that the older Llama models would shy away from. Plus, Meta claims Llama 4 is "dramatically more balanced" in what it refuses to answer.

"You can count on Llama 4 to give you helpful, factual answers without judgment," a Meta spokesperson told TechCrunch. "We're working to make Llama more responsive to a variety of viewpoints and less likely to favor some over others."

This move comes as some of President Donald Trump's allies, like Elon Musk and David Sacks, have accused AI chatbots of being too "woke" and censoring conservative views. Sacks has specifically called out OpenAI's ChatGPT for being biased and untruthful about political topics.

The truth is, getting rid of bias in AI is a tough nut to crack. Even Musk's own AI company, xAI, has had trouble making a chatbot that doesn't lean one way or another politically.

Despite this, companies like OpenAI have been adjusting their models to answer more questions, especially those about controversial subjects.