Meta’s Llama 4 Scout & Maverick: Open-Source AI Models You Can Use Now

Meta launches Llama 4 Scout and Maverick—efficient, open-source multimodal AI models that outperform GPT-4o and run on a single GPU. Free to download now.

Meta Unveils Llama 4 Scout and Maverick

Meta has officially introduced two powerful AI models — Llama 4 Scout and Llama 4 Maverick, making them the first open-source releases from its Llama 4 family. Announced by Mark Zuckerberg and available across Meta platforms and Meta.ai, these models are already being hailed as next-gen multimodal tools.
Let’s look at what makes Llama 4 Scout stand out from the crowd.

Llama 4 Scout: Small Size, Massive Power

Llama 4 Scout packs a punch with:

• 17 billion active parameters
• 16 experts using a Mixture-of-Experts (MoE) framework
• 10 million token context window
• Compatible with just a single Nvidia H100 GPU

Despite its smaller footprint, Scout surpasses GPT-4o, Gemini Flash 2, and DeepSeek V3 in core performance benchmarks.

But how does Maverick compare in terms of multimodal versatility? Let’s dive in.

Llama 4 Maverick: Multimodal and Mighty

Maverick, like Scout, uses 17B active parameters but boosts expert capacity to 128, enhancing its understanding across text, images, video, and audio. It’s tailored for high-performance reasoning and coding, achieving more with fewer parameters than many competitors.

What makes both models so efficient under the hood is their shared MoE architecture. Here's why it matters.

The MoE Advantage: Efficiency by Design

The Mixture-of-Experts design activates only the experts required for each task, reducing energy and computation without compromising output. This architecture enables both models to deliver real-time results even in low-resource environments.
Now, where and how can you access these models? Let’s explore availability.

Where to Download + What’s Coming Next

Both Llama 4 Scout and Maverick are available for free on Meta and Hugging Face. Scout is already integrated with WhatsApp, Messenger, and Instagram Direct. Up next: Llama 4 Reasoning, followed by Llama 4 Behemoth — a 2 trillion-parameter model currently in training.
This signals just the beginning of Meta’s open-source AI evolution.