Meta picked a Saturday to drop Llama 4, its latest AI posse, because nothing screams “weekend vibes” like a tech bombshell. Say hello to Llama 4 Scout, Llama 4 Maverick, and the still-baking Llama 4 Behemoth.
These models guzzled heaps of unlabeled text, images, and videos to gain what Meta calls “broad visual understanding,” ready to tackle chats, code, and whatever else you throw their way.
China’s DeepSeek lab lit a fire under Meta’s backside with its open models, R1 and V3, which reportedly smoked the last Llama crew.
Word is, Meta turned into a frantic war-room sitcom, reverse-engineering DeepSeek’s budget-friendly magic to birth Llama 4. Now, Scout and Maverick are strutting their stuff on Llama.com and Hugging Face, while Behemoth’s still pumping iron.
Meta AI, the assistant sprinkled across WhatsApp, Messenger, and Instagram, is already rocking Llama 4 in 40 countries. Multimodal flair, like mixing images with text, is a U.S.-only treat for now, and only in English.
Here’s the rub: not everyone gets a ticket. EU folks are flat-out barred from using or sharing Llama 4, probably Meta’s way of flipping off the region’s strict AI and privacy rules.
Got over 700 million monthly users? You’ll need to grovel for a special license, and Meta can shrug you off for kicks. Still, they’re hyping this as the start of a Llama revolution. We’ll believe it when we see it.
What’s cooking inside?
Llama 4 uses a “mixture of experts” (MoE) trick, splitting tasks among mini-models for leaner, meaner performance. Maverick, the chatty jack-of-all-trades, packs 400 billion total parameters but runs on just 17 billion active ones across 128 experts.
Scout, the document-crunching nerd, has 109 billion total parameters, 17 billion active, and a ridiculous 10-million-token context window, enough to digest a library and still burp out insights.
Behemoth’s the muscle, with 288 billion active parameters and nearly two trillion total, though it’s not flexing yet.
Maverick beats OpenAI’s GPT-4o and Google’s Gemini 2.0 in coding and creative writing, per Meta’s brag sheet, but it’s no match for champs like GPT-4.5 or Claude 3.7 Sonnet.
Scout shines at summarizing tomes or debugging epic code, while Behemoth’s future scores hint it’ll crush GPT-4.5 in math and STEM. None of them are “reasoning” models like OpenAI’s o1, though, so don’t expect them to ponder life’s mysteries.
Meta dialed back Llama 4’s prim-and-proper act. It’ll wade into spicy political and social debates where older models clammed up, aiming for “helpful, factual responses without judgment.”
They claim it’s “dramatically more balanced,” a jab at the “woke AI” gripes from Elon Musk’s circle. Irony alert: Musk’s xAI hasn’t exactly nailed impartiality either.
Hardware-wise, Scout’s a lightweight, sipping juice on one Nvidia H100 GPU. Maverick demands a beefy H100 DGX setup. Behemoth? Better rent a data center.
Meta’s teasing more Llama 4 surprises ahead, clearly itching to dominate the AI game. For now, EU coders and bias hawks might just have to watch from the sidelines, popcorn in hand.
Follow TechBSB For More…
