Meta Unveils Llama 4 Series: New Scout, Maverick & Behemoth AI Models

Meta has significantly advanced its AI capabilities with the launch of its Llama 4 series, introducing two new powerful and versatile models, Llama 4 Scout and Llama 4 Maverick. These models feature advanced architecture and are designed for different strengths, signaling Meta's continued push in the competitive AI landscape.

Introducing Llama 4 Scout & Maverick

Both Scout and Maverick share some core characteristics: they are multimodal (capable of processing text, images, video, audio), feature a Mixture-of-Experts (MoE) architecture for efficiency, and have 17 billion active parameters. However, they are optimized differently:

• Llama 4 Scout: Built with 16 experts and optimized for performance on a single Nvidia H100 GPU. It boasts an impressive 10 million token context window and reportedly outperforms several rival models in benchmarks.

• Llama 4 Maverick: Features a larger number of experts (128) and is focused more on efficiency and cost-effectiveness, while still delivering strong results in reasoning and coding tasks.

Technical Advancements & Training

The Llama 4 series represents a step forward in model training and capabilities. Meta highlighted:

• Diverse Training Data: Trained on varied datasets including text, images, and video.

• Efficiency Techniques: Employed methods like MetaP and FP8 precision during training to enhance quality and efficiency.

• Multilingual Support: Capable of supporting over 200 languages.

Availability & Integration

Developers can access both Llama 4 Scout and Maverick via platforms like Hugging Face and the official Llama website. Furthermore, Meta plans to integrate these advanced AI capabilities across its own platforms, including WhatsApp, Messenger, and Instagram Direct, suggesting users will soon experience these models powering features within the apps they use daily.

What About Behemoth?

Meta also offered a glimpse of an even larger model still in training: Llama 4 Behemoth. This model reportedly has 288 billion parameters and is already showing strong performance, particularly on STEM-focused benchmarks, indicating Meta's ambition in building frontier-level AI systems.

About Meta's AI Strategy

This launch is part of Meta's broader, significant investment in AI. The company is pouring billions into data center infrastructure (including a new $1B center in Wisconsin and an $837M one elsewhere in the US), developing in-house AI chips to potentially reduce reliance on Nvidia, and pushing research via its FAIR group (Fundamental AI Research), despite recent high-level departures like the head of AI research. Meta is also reportedly considering a standalone AI app, potentially with subscription models.

Looking Ahead

The release of the Llama 4 series, particularly the accessible Scout and Maverick models, provides developers and businesses (including those in MENA) with powerful new tools. The efficiency focus of Maverick could be particularly appealing. More details and updates are expected at Meta's LlamaCon event on April 29th. This launch heats up the AI race, putting Meta's latest offerings in direct comparison with models from OpenAI, Google, Anthropic, and others.

Source: TechInAsia