Microsoft Officially Announces Open Source Magma Multimodal AI , Easily Holds Web Pages, Bots

The Microsoft Open-source Multi-Modular AI Basic Model Magma, which understands multi-modular input into the environment and is relevant to reality, has been accepted by the CVPR to support web-based navigation and robotic operations; Magma has innovatively proposed two labels: Set-of-Mark, which provides a high-level “care mark” for key audiences, and Trade-of-Mark, which captures time-series changes in action; the model is pre-trained by more than 39 million samples, using the ConvNeXt-XXL visual network and the Llama-3-8B language model, with a majority of the team being Chinese and Yang Jianwei, a senior Microsoft researcher。