Meta Drops Llama 4: A Fresh Batch of Flagship AI Models
Menlo Park, CA – April 6, 2025 – Meta has unleashed Llama 4, its latest suite of flagship AI models, signaling a bold step forward in its mission to dominate the open-source AI frontier. Rolled out over the weekend, Llama 4 introduces two models available now—Llama 4 Scout and Llama 4 Maverick—while two more, Llama 4 Behemoth and Llama 4 Reasoning, are poised to join the lineup soon. This release cements Meta’s ambition to outpace rivals like OpenAI and Google with cutting-edge tech that’s free for most to use.
Llama 4 Scout and Maverick are live today, downloadable from Llama.com and Hugging Face, and already woven into Meta AI across its apps—WhatsApp, Instagram, Messenger—and the web. Scout, a lightweight 17-billion-parameter model with 16 experts, is a speed demon, running on a single Nvidia H100 GPU. Its standout feature? A jaw-dropping 10-million-token context window, letting it chew through massive datasets—like entire novels or sprawling code repos—in one gulp. It’s tuned for tasks like summarizing stacks of documents or parsing user habits with eerie precision. Maverick, also 17 billion active parameters but with 128 experts and 400 billion total, flexes multimodal muscle, blending text and image inputs for chatbots and assistants. Meta says it beats GPT-4o and Gemini 2.0 Flash in coding, reasoning, and multilingual tests, nipping at DeepSeek V3.1’s heels.
These models mark Llama’s first dive into mixture-of-experts (MoE) territory, a design that keeps compute costs low while boosting performance. They’re distilled from the still-cooking Llama 4 Behemoth, a 288-billion-parameter titan with nearly 2 trillion total, which Meta’s Mark Zuckerberg touts as “the best base model out there” based on internal STEM benchmarks topping GPT-4.5 and Claude 3.7 Sonnet. Behemoth’s full debut is imminent, while Llama 4 Reasoning, a logic-focused fourth model, is teased for May, with more to spill at LlamaCon on April 29.
Meta’s sticking to its open-source guns—mostly. The EU’s sidelined by regulatory red tape, and big players with over 700 million monthly users need a hall pass to join the party. With a $65 billion AI budget this year, Zuckerberg’s betting Llama 4 will hook developers and startups worldwide. X chatter is electric, hyping Scout’s context length and Maverick’s all-around prowess. As the AI arms race accelerates, Llama 4’s arrival could redraw the map—stay tuned.