Llama 4 Multimodal AI: Training, Capabilities and Performance

In the race for which AI tool is the best, how can we forget Meta Platforms, Inc., which launched Llama (Large Language Model Meta AI) in 2023? Recently, on April 05, 2025, Meta added +1 new model in its Llama iteration, and that is ‘The Llama 4 herd. ’

Most of you must be wondering what this Llama 4 herd is and for what type of target audience it has been made. 

For instance, OpenAI’s OpenAI Academy has been made for developers, educators, students, and professionals. The Llama 4 herd targets researchers, developers, educators, AI enthusiasts, and more, catering to different types of audiences.

By rolling out its latest models in the LLaMA lineup, named LLaMA 4 Scout and LLaMA 4 Maverick, these are the first open-weight and multimodal models in this series. They come with a long context length and use a mixture-of-experts (MoE) approach.

Meta also gave us a sneak peek at LLaMA 4 Behemoth. This model is one of the smartest large language models around and will help train the new models.

What is Llama 4 Herd? The Beginning of a New Era

“The Llama 4 herd: The beginning of a new era of natively multimodal AI innovation.” What do you understand by reading this line? 

The Llama 4 herd is a multimodal AI, which means it has the ability to understand and work with multiple types of input, like text, images, audio, or video, at the same time. 

Meta introduced the Llama 1, Llama 2, and Llama 3 series. Now, Llama 4 includes Behemoth, Maverick, and Scout. These new LLaMA 4 models mark a major step forward for the LLaMA ecosystem. Meta has designed two efficient models in this series:

LLaMA 4 Scout: A 17-billion active parameter model with 16 experts. It can run on a single H100 GPU using Int4 quantisation.

LLaMA 4 Maverick: Also a 17-billion active parameter model, but with 128 experts. It fits on a single H100 host.

Meta also developed a powerful teacher model called LLaMA 4 Behemoth. This model outperforms GPT-4.5, Claude Sonnet 3.7, and Gemini 2.0 Pro on STEM-heavy benchmarks like MATH-500 and GPQA Diamond.

Although LLaMA 4 Behemoth is still in training and not available yet, Meta plans to share more technical insights soon.

Claude 3.5 Sonnet vs GPT-4o vs Gemini 1.5: Which is the Most Powerful AI Model?

Llama 4 Herd: Meet Behemoth, Maverick and Scout

Last week, Meta Platforms revealed the first models in the LLaMA 4 family. These are meant to create more personalised and versatile AI experiences.

LLaMA 4 Scout:

• Has 17 billion active parameters and uses 16 experts.
• One of the most powerful multimodal models in its class.
• Fits on a single NVIDIA H100 GPU (using Int4 quantisation).
• Offers a massive 10 million token context window.
• Outperforms models like Gemma 3, Gemini 2.0 Flash-Lite, and Mistral 3.1 on multiple benchmarks.

LLaMA 4 Maverick:

• Also has 17 billion active parameters but with 128 experts.
• Beats GPT-4o and Gemini 2.0 Flash on a wide range of popular benchmarks.
• Matches DeepSeek v3 in reasoning and coding while using less than half the active parameters.
• Delivers an excellent performance-to-cost ratio.
• Its experimental chat model scored 1417 ELO on LMArena.

LLaMA 4 Behemoth

• Both models were improved through distillation from LLaMA 4 Behemoth, Meta’s most powerful model yet, which has 288 billion active parameters and 16 experts.
• Outperforms GPT-4.5, Claude Sonnet 3.7, and Gemini 2.0 Pro on STEM benchmarks.
• Still in training, but Meta plans to share more technical insights soon.

Availability:

You can download LLaMA 4 Scout and Maverick from llama.com or Hugging Face. Also, you can try Meta AI powered by LLaMA 4 in WhatsApp, Messenger, Instagram Direct, and on the web.

Llama 3.1 vs GPT 4 vs Mixtral 8x22B vs Claude 3.5: Which is Best LLM Model?