Meta Unveils Two Llama 4 AI Models
Hey, did you hear? Meta just rolled out Llama 4, their latest set of AI models that are now jazzing up the Meta AI assistant on the web, WhatsApp, Messenger, and Instagram. They've launched two new models you can snag from Meta or Hugging Face: Llama 4 Scout and Llama 4 Maverick. Scout's a nifty little model that can squeeze into a single Nvidia H100 GPU, while Maverick's more like the big guns, GPT-4o and Gemini 2.0 Flash. And guess what? They're still cooking up Llama 4 Behemoth, which Mark Zuckerberg says is gonna be "the highest performing base model in the world."
So, about Llama 4 Scout—it's got this huge 10-million-token context window, which is like its working memory, and it's knocking out Google's Gemma 3 and Gemini 2.0 Flash-Lite, plus the open-source Mistral 3.1, in all sorts of benchmark tests. And the best part? It still fits in that single Nvidia H100 GPU. Maverick's not too shabby either, holding its own against OpenAI's GPT-4o and Google's Gemini 2.0 Flash, and it's giving DeepSeek-V3 a run for its money in coding and reasoning tasks with way fewer parameters.
Image: Meta
Now, Llama 4 Behemoth's a beast with 288 billion active parameters and a whopping 2 trillion in total. It's not out yet, but Meta's already boasting that it'll outshine GPT-4.5 and Claude Sonnet 3.7 in STEM benchmarks.
For Llama 4, Meta's switched things up with a "mixture of experts" (MoE) architecture, which is pretty smart because it only uses the parts of the model needed for each task, saving resources. They're gonna spill more beans about their future AI plans at LlamaCon on April 29th, so keep an eye out for that.
Just like before, Meta's calling Llama 4 "open-source," but there's been some pushback on that. The Llama 4 license has some rules, like if you're a big commercial outfit with over 700 million monthly active users, you gotta ask Meta for permission before using it. The Open Source Initiative wasn't too happy about that back in 2023, saying it doesn't really fit the "open-source" bill.
Related article
WordPress.com now allows AI agents to write and publish posts, plus more
WordPress.com, the popular web hosting and publishing platform, is now embracing AI agents—a move that could reshape the look and feel of the web. The company announced Friday that it will allow AI agents to draft, edit, and publish content on custom
Meta AI now responds to buyer messages on Facebook Marketplace
Facebook Marketplace introduces new Meta AI features, including automated replies to buyer inquiries, the company announced Thursday. The platform also leverages AI to accelerate item listings, summarize seller profiles, and now lets sellers offer sh
Meta signs deal for millions of Amazon AI CPUs
Amazon has secured a significant partnership with Meta, once again relying on its own custom-designed chips. Meta has agreed to deploy millions of AWS Graviton chips to meet its expanding AI demands, Amazon confirmed on Friday.Note that AWS Graviton
Related Special Topic Recommendations
Comments (54)
0/500
Llama 4の発表、特にScoutとMaverickの分業が気になりますね。Scoutが軽量で速いのは分かるけど、実際のアシスタント体験でどれくらい差が出るのかな?個人的には、こうしたモデルの細分化が進むと、結局は開発者や企業がコストと性能のトレードオフで迷うだけのような気も… 🤔 でも、とりあえずHugging Faceで触ってみようかな。
Interessant, dass Meta jetzt gleich zwei Versionen von Llama 4 rausbringt! 🤔 Scout klingt nach einem guten Einstiegsmodell, aber ich frage mich, ob Maverick wirklich so viel besser ist... Vielleicht sollten sie lieber einen echt krassen Fokus statt viele halbgare Modelle entwickeln. Aber trotzdem nice, dass die Modelle auf Hugging Face verfügbar sind!
Wow, Meta's Llama 4 sounds like a game-changer! Scout and Maverick? Cool names, but I wonder if they’ll outsmart my cat at hide-and-seek. 😺 Curious to see how these jazz up WhatsApp!
Wow, Llama 4 sounds like a game-changer for Meta AI! I'm curious how Scout and Maverick stack up against other models—any cool new tricks up their sleeves? 😎
Probé el Llama 4 de Meta en WhatsApp y ¡es genial! El modelo Scout es muy útil para respuestas rápidas, pero Maverick parece demasiado atrevido. Sin embargo, es una buena actualización para el AI de Meta. Tal vez la próxima vez puedan hacer que Maverick sea menos 'maverick'? 😂
Hey, did you hear? Meta just rolled out Llama 4, their latest set of AI models that are now jazzing up the Meta AI assistant on the web, WhatsApp, Messenger, and Instagram. They've launched two new models you can snag from Meta or Hugging Face: Llama 4 Scout and Llama 4 Maverick. Scout's a nifty little model that can squeeze into a single Nvidia H100 GPU, while Maverick's more like the big guns, GPT-4o and Gemini 2.0 Flash. And guess what? They're still cooking up Llama 4 Behemoth, which Mark Zuckerberg says is gonna be "the highest performing base model in the world."
So, about Llama 4 Scout—it's got this huge 10-million-token context window, which is like its working memory, and it's knocking out Google's Gemma 3 and Gemini 2.0 Flash-Lite, plus the open-source Mistral 3.1, in all sorts of benchmark tests. And the best part? It still fits in that single Nvidia H100 GPU. Maverick's not too shabby either, holding its own against OpenAI's GPT-4o and Google's Gemini 2.0 Flash, and it's giving DeepSeek-V3 a run for its money in coding and reasoning tasks with way fewer parameters.
Image: Meta
Now, Llama 4 Behemoth's a beast with 288 billion active parameters and a whopping 2 trillion in total. It's not out yet, but Meta's already boasting that it'll outshine GPT-4.5 and Claude Sonnet 3.7 in STEM benchmarks.
For Llama 4, Meta's switched things up with a "mixture of experts" (MoE) architecture, which is pretty smart because it only uses the parts of the model needed for each task, saving resources. They're gonna spill more beans about their future AI plans at LlamaCon on April 29th, so keep an eye out for that.
Just like before, Meta's calling Llama 4 "open-source," but there's been some pushback on that. The Llama 4 license has some rules, like if you're a big commercial outfit with over 700 million monthly active users, you gotta ask Meta for permission before using it. The Open Source Initiative wasn't too happy about that back in 2023, saying it doesn't really fit the "open-source" bill.
WordPress.com now allows AI agents to write and publish posts, plus more
WordPress.com, the popular web hosting and publishing platform, is now embracing AI agents—a move that could reshape the look and feel of the web. The company announced Friday that it will allow AI agents to draft, edit, and publish content on custom
Meta AI now responds to buyer messages on Facebook Marketplace
Facebook Marketplace introduces new Meta AI features, including automated replies to buyer inquiries, the company announced Thursday. The platform also leverages AI to accelerate item listings, summarize seller profiles, and now lets sellers offer sh
Meta signs deal for millions of Amazon AI CPUs
Amazon has secured a significant partnership with Meta, once again relying on its own custom-designed chips. Meta has agreed to deploy millions of AWS Graviton chips to meet its expanding AI demands, Amazon confirmed on Friday.Note that AWS Graviton
Llama 4の発表、特にScoutとMaverickの分業が気になりますね。Scoutが軽量で速いのは分かるけど、実際のアシスタント体験でどれくらい差が出るのかな?個人的には、こうしたモデルの細分化が進むと、結局は開発者や企業がコストと性能のトレードオフで迷うだけのような気も… 🤔 でも、とりあえずHugging Faceで触ってみようかな。
Interessant, dass Meta jetzt gleich zwei Versionen von Llama 4 rausbringt! 🤔 Scout klingt nach einem guten Einstiegsmodell, aber ich frage mich, ob Maverick wirklich so viel besser ist... Vielleicht sollten sie lieber einen echt krassen Fokus statt viele halbgare Modelle entwickeln. Aber trotzdem nice, dass die Modelle auf Hugging Face verfügbar sind!
Wow, Meta's Llama 4 sounds like a game-changer! Scout and Maverick? Cool names, but I wonder if they’ll outsmart my cat at hide-and-seek. 😺 Curious to see how these jazz up WhatsApp!
Wow, Llama 4 sounds like a game-changer for Meta AI! I'm curious how Scout and Maverick stack up against other models—any cool new tricks up their sleeves? 😎
Probé el Llama 4 de Meta en WhatsApp y ¡es genial! El modelo Scout es muy útil para respuestas rápidas, pero Maverick parece demasiado atrevido. Sin embargo, es una buena actualización para el AI de Meta. Tal vez la próxima vez puedan hacer que Maverick sea menos 'maverick'? 😂





Home






