Llama 4: Meta Claims Latest Open-Source AI Model Beats Google, OpenAI Models

In its push into the artificial intelligence space, Meta has released its latest set of AI models, known as Llama 4, which power the Meta AI assistant integrated across web, Instagram Direct, WhatsApp and Messenger. 

Known as Llama 4 Scout and Llama 4 Maverick, they are the first two open-source AI models in the Llama 4 family, and two more models are anticipated to be released later.

"Llama 4 is a milestone for Meta AI and for open source. For the first time, the best small, mid-sized and potentially soon, frontier models will be open source," Meta Chief Executive Officer Mark Zuckerberg said in an Instagram Reel. 

The Llama 4 Scout — the smaller multimodal model — can operate on a single GPU, such the Nvidia H100, and has a context length of 10 million tokens with 17 billion parameters and 16 experts. 

Meta claims that Llama 4 Scout surpasses Google's Gemma 3 and Gemini 2.0 Flash-Lite models, as well as the open-source Mistral 3.1 model, "across a broad range of widely reported benchmarks" and still fits in a single Nvidia H100 GPU. 

Meta also claims that its larger Llama 4 Maverick performs similarly to DeepSeek-V3 in coding and reasoning with "less than half the active parameters" and surpasses OpenAI's GPT-4o and Google's Gemini 2.0 Flash.

With 128 experts and 17 billion parameters, the Maverick is designed to be a general-purpose model for a variety of assistant-style use cases. According to Meta, it is a dependable “workhorse” that can manage digital activities like reasoning and chat.

The Llama 4 Behemoth was also previewed, and according to Meta, it is "one of the smartest LLMs in the world and our most powerful yet to serve as a teacher for our new models", with 288 billion active parameters a total of 2 trillion parameters. 

The Llama 4 Scout and Llama 4 Maverick models can be downloaded from Meta's Llama website or Hugging Face.