Meta on Saturday ... similar to the DeepSeek v3 model, using less than half the resources to do so. The models are built on a new “mixture of experts” (MoE) architecture.

Meta has also previewed Llama 4 Behemoth, the largest AI model in the family so far, with 288 billion active parameters.

The Llama 4 series is the first to use a “mixture of experts (MoE) architecture,” where only a few parts of the neural ...

Revelations about Meta's secret tests highlight how valuable training data is for AI model development, and suggest a way for ...

In a statement, Meta said, "People's interactions with Meta AI – like questions and queries – will also be used to train and improve our models. This training, which follows the successful ...

Meta Platforms has rolled out the newest iteration of its large language model—Llama 4—featuring two variants: Llama 4 Scout and Llama 4 Maverick. According to Meta, Llama 4 is a multimodal AI ...

Meta announced on Monday that it’s going to train its AI models on public content, such as posts and comments on Facebook and Instagram, in the EU after previously pausing its plans to do so in ...

Meta decided to pause the launch of its AI models in Europe last June after Ireland's Data Protection Commission (DPC) told the company to delay its plan to harness data from social media posts.

Meta will train its artificial intelligence (AI) models with its European users' public content and conversations with the Meta AI chatbot, the firm said on Monday. The decision represents a major ...

For Llama 4, Meta adopted a “mixture of experts” (MoE) architecture—a setup that helps save computing power by activating only the specific parts of the model needed for each task.