Meta releases open-weight Llama 4 AI models to rival DeepSeek, Google Gemma

Meta has introduced Llama 4, its newest series of open-weight AI models, which will now improve the abilities of Meta AI across services like WhatsApp, Instagram, and Messenger. Two of the new models, Llama 4 Scout and Llama 4 Maverick, are available for download via Meta’s official Llama website and Hugging Face. The company has also previewed Llama 4 Behemoth, a teacher model designed to train and guide other Llama 4 models.

 

Open-weight AI models are those with publicly available trained weights, allowing users to run them locally without relying on cloud APIs. While accessible, they may still have licensing restrictions on changes or commercial use.

Llama 4 AI models: Details

During the announcement, Meta described Llama 4 as its most advanced set of AI models yet, built to support the broader Llama ecosystem. These models aim to help developers and businesses create more tailored and multimodal AI-driven experiences.

 

  Llama 4 Scout and Llama 4 Maverick are Meta’s first open-weight AI models that are natively multimodal and built on a mixture-of-experts (MoE) architecture. Below are the highlights of each model:

 

Llama 4 Scout

 

Llama 4 Scout is designed as a smaller model, using 17 billion active parameters and 16 experts. It supports a context window of up to 10 million tokens and can run on a single Nvidia H100 GPU (graphic processing unit). According to Meta, Scout is better than Google’s Gemma 3, Gemini 2.0 Flash-Lite, and Mistral 3.1 on various benchmark tests.

 

Llama 4 Maverick

 

Maverick is the more advanced offering in the set, positioned as a competitor to OpenAI’s GPT-4o and Gemini 2.0 Flash. Meta claims that Maverick not only performs better than these models in several benchmarks but also rivals DeepSeek v3 in reasoning and coding tasks—all while using less than half the active parameters.

 

Using 17 billion active parameters and 128 experts, Llama 4 Maverick is optimised to run on a single Nvidia H100 host.

 

Llama 4 Behemoth

 

The Behemoth model acts as a teacher model and is the most powerful variant within the Llama 4 family. This multimodal MoE model features 288 billion active parameters, uses 16 experts, and has nearly two trillion total parameters. According to Meta, it delivers leading performance in maths, multilingual processing, and image-related tasks among non-reasoning models, making it ideal for training its smaller counterparts.

 

Although still being trained, Meta says that Llama 4 Behemoth is better than OpenAI’s GPT-4.5, Claude Sonnet 3.7, and Google Gemini 2.0 Pro on several STEM-focused evaluations.

Llama 4 AI models: Availability

Llama 4 Scout and Llama 4 Maverick are now available for download from Meta’s official Llama page and on Hugging Face. Furthermore, Meta AI powered by Llama 4 is gradually becoming available across WhatsApp, Messenger, Instagram Direct, and web platforms.