悟道

Overview

Wudao is an ultra-large-scale intelligent model system launched by the Beijing Academy of Artificial Intelligence (BAAI). Wudao 2.0 was released in June 2021 and is one of the earlier important research achievements in China oriented toward large models. The project is positioned for multimodal and large-scale pre-trained model research, combining Chinese and English text, images, and other data for training to explore stronger general representation and generation capabilities.

Based on public information, Wudao 2.0 was trained on a large-scale, high-quality dataset, with total training data of about 4.9TB, including Chinese text, Chinese image-text data, and English text data. Its underlying training adopted the MoE (Mixture of Experts) architecture based on the open-source FastMoE to improve the training efficiency and inference capabilities of ultra-large models.

As a model system with strong research attributes, Wudao is more appropriately understood as a foundational capability platform for AI research, large-model training, and multimodal modeling, rather than a consumer-grade application for direct use by ordinary users.

Main Features

• Ultra-large-scale pre-trained model research

  • Conducts foundational model exploration for language understanding, generation, and general intelligence.

• Chinese and English multi-source data training

  • Uses Chinese text, English text, and Chinese image-text data for joint training, supporting richer semantic modeling capabilities.

• Multimodal capability exploration

  • Covers not only text, but also combines image and text data, making it suitable for multimodal pre-training research scenarios.

• Training based on the MoE architecture

  • Uses a mixture-of-experts mechanism, assigning different tasks or inputs to more suitable "expert models" for processing, in order to improve training and inference efficiency.

• Built on the FastMoE open-source system

  • Uses FastMoE for model training, supporting large-scale parallel training and the implementation of expert routing mechanisms.

• Serves research and development scenarios

  • Suitable for developers and research teams focused on large models, pre-training, model architecture design, and research into the underlying capabilities of artificial intelligence.

Product Pricing

The public page currently does not provide clear commercial pricing information.

Based on the official website and existing materials, Wudao is more oriented toward the presentation of research projects and model system achievements. Whether it offers an open API, commercial access methods, usage fees, or enterprise cooperation plans should be subject to the latest official statement from the Beijing Academy of Artificial Intelligence.

Frequently Asked Questions

Who is Wudao suitable for?

It is more suitable for artificial intelligence researchers, university laboratories, algorithm engineers, and development teams focused on the underlying technologies of large models.

Is Wudao a chatbot?

Public information shows that Wudao is essentially an ultra-large-scale pre-trained model system and should not be simply equated with a chatbot product for end users.

What types of training data does Wudao have?

Publicly available information includes:

• 1.2TB of Chinese text data
• 2.5TB of Chinese image-text data
• 1.2TB of English text data

What are Wudao’s technical characteristics?

Its representative feature is the use of the MoE (Mixture of Experts) architecture, and it is trained based on FastMoE, enabling the model to dynamically call more suitable expert modules according to the input content.