Meta Unveils Llama 4: Scout, Maverick, & What's Next

Meta has officially kicked off the next generation of its open-weight AI family with the launch of Llama 4, introducing two powerful new models: Scout and Maverick—now live inside Meta AI across WhatsApp, Messenger, and Instagram. With two more models on the way, this marks a major leap in Meta’s push toward efficient, multimodal AI.

“Today is the start of a new era of natively multimodal AI innovation.”
— @AIatMeta, April 5, 2025

Meet the New Llama 4 Models

These are the two elements you need to know about.

Maverick: The Multimodal Workhorse

• 400B total parameters, with only 17B active via MoE (mixture-of-experts) for maximum efficiency.

• Optimized for general-purpose assistant tasks, from search to scheduling.

• Integrated across Meta’s suite of apps and currently ranked #2 on LM Arena.

Scout: The Specialist

• Designed for summarization and code reasoning.

• Supports an eye-popping 10 million token context window.

• Lightweight enough to run on a single Nvidia H100 GPU.

Both models use MoE architecture to deliver high performance at lower compute costs, supporting native multimodal tasks like image + text interactions.

On the Horizon: Behemoth & Reasoning

Meta also teased two upcoming Llama 4 variants:

• Llama 4 Behemoth: A still-training monster model with 288B active parameters, reportedly outperforming GPT-4.5 and Claude 3.7 Sonnet on STEM tasks.

• Llama 4 Reasoning: Launching within weeks, this model is aimed at true reasoning capabilities, including advanced fact-checking.

While Behemoth is positioned as a next-gen base model, neither it nor the current models qualify as fully “reasoning” AIs just yet—a benchmark Meta appears determined to hit.

Access (and Restrictions)

Meta continues its tradition of open-weight releases—with a catch. Due to ongoing regulatory uncertainty under the EU AI Act, EU-based developers are barred from using Llama 4 models for now.

This geo-fencing highlights the growing friction between rapid AI development and evolving global regulation.

Benchmark Drama: The Maverick Mismatch

Despite strong LM Arena rankings, Maverick has come under scrutiny. Researchers discovered that the version benchmarked is not the same as the one released publicly. The LM Arena model is a “chat-optimized” variant, while the publicly released version features excessive emojis and overly verbose responses.

This discrepancy raises serious questions about benchmark integrity—are some models fine-tuned just to ace the leaderboard, while everyday users and developers get something different?

Hot Bleat: Llama 4 is More Than  Model Fam

Meta’s Llama 4 launch shows both ambition and a bit of turbulence. The models are fast, scalable, and multimodal—but early feedback underscores the importance of transparency and consistency. With two more models on the way, including one focused on reasoning, the race for truly intelligent AI assistants is heating up fast.

Stay tuned—Llama 4 isn’t just a model family. It’s Meta’s roadmap for the future of open, multimodal AI.