Meta has officially launched Llama 4, the next-generation suite of open-source AI models, marking a major milestone in its growing AI ecosystem. Announced on April 5 by CEO Mark Zuckerberg, Llama 4 introduces multimodal capabilities, enhanced performance, and a unique Mixture-of-Experts (MoE) architecture aimed at improving efficiency and scalability.

Llama 4 Model Lineup: Scout and Maverick

At launch, Meta introduced two distinct models:

Call 4 scout

17 billion parameters

16 experts

10 million token context window

Runs on a single GPU

Designed for lightweight applications

Call 4 maverick

17 billion parameters

128 experts

General-purpose assistant model

Runs on a single host

Targets coding, reasoning, multilingual, long-context, and image tasks

Meta claims it outperforms GPT-4o and Gemini 2.0 Flash

Both models are available via the Llama website and Hugging Face, and will power Meta AI across WhatsApp, Instagram, Messenger, and web platforms.

Llama 4 Behemoth and Future Models

Meta also previewed Llama 4 Behemoth, a model currently in training:

288 billion active parameters

16 experts with nearly 2 trillion total parameters

Meta claims Behemoth will surpass GPT-4.5, Claude Sonnet 3.7, and Gemini 2.0 Pro in STEM benchmarks once fully developed.

Another upcoming variant is Llama 4 Reasoning, aimed at advanced cognitive tasks. Meta plans to reveal more details in the coming month.

Mixture-of-Experts (MoE): Efficiency & Intelligence

Llama 4 models adopt the MoE architecture, where only a subset of domain-specific experts are activated for each task. This allows:

Reduced training and inference costs

Faster processing and higher efficiency

Domain-specialized performance (e.g., programming, biology, physics)

MoE is gaining attention globally, especially after Chinese lab DeepSeek showcased its success with the approach.

Competitive Edge in the AI Race

Meta’s Llama 4 arrives amidst a fierce AI race alongside Google’s Gemini 2.0 and OpenAI’s GPT-4o. Zuckerberg stated:
“Open-source AI is becoming the dominant model. With Llama 4, that shift is clearly underway.”

Despite DeepSeek’s growing influence, Meta executives remain confident in Llama’s position as a leader in open-source AI.

Read More: Meta Launches Llama 4 with Multimodal Capabilities and MoE Architecture