Facebook founder creates new AI model.

Facebook's parent company has just announced two new models in its Llama 4 collection. These are the same models currently used for the Meta AI assistant on the web, as well as in WhatsApp, Messenger, and Instagram, helping people create more personalized multimedia experiences.

Two new models, now available for download from Meta or Hugging Face, are the Llama 4 Scout and the Llama 4 Maverick. Meta says it is still in the process of training the Llama 4 Behemoth, the massive model that CEO Mark Zuckerberg calls “ the world’s highest-performing platform model.”

According to Meta, Llama 4 Scout has a context window of up to 10 million tokens, many times more than the model with the highest number of tokens currently, Gemini. This number represents the working memory of an AI model.

Meta describes the model as superior to Google's Gemma 3 and Gemini 2.0 Flash-Lite, as well as Mistral 3.1, "on a number of widely published benchmarks," while still being able to "run on a single Nvidia H100 GPU."

For Llama 4 Maverick, the company made similar performance claims when comparing it to OpenAI's GPT-4o and Google's Gemini 2.0 Flash. Maverick's programming and inference tasks are comparable to DeepSeek-V3, even though it only uses less than half the parameters.

Figures for the new models in the Llama 4 collection. Photo: Meta.

Meanwhile, Behemoth boasts 288 billion active parameters out of a total of 2 trillion. Although not yet released, Meta states that Behemoth has the potential to outperform its competitors (specifically GPT-4.5 and Claude Sonnet 3.7) “on several STEM metrics.”

Meta announced that they have switched to a “mixture of experts” (MoE) architecture, the same architecture that made DeepSeek so successful. This approach saves resources by only activating the parts of the model necessary for each specific task. More details about the product plan will be announced at the LlamaCon conference on April 29th.

Despite criticism regarding its licensing restrictions, Meta continues to label Llama 4 as "open source," similar to its previous models. For example, the Llama 4 license requires commercial organizations with more than 700 million monthly active users to obtain permission from Meta before using the model.

According to the Open Source Initiative in 2023, such regulations "have rendered the model no longer classified as open source."