News
ReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior ReleaseReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior Release
EveryLLM
imgWu Dao

Wu Dao

BAAI's trillion-parameter multimodal model family, announced 2021

Wu Dao is a multimodal AI model family from the Beijing Academy of Artificial Intelligence, first announced on 11 January 2021 with a larger 2.0 version on 31 May 2021. Its 1.75 trillion parameters and mixture-of-experts design drew comparisons with GPT-3, and it handles text, images and protein structure tasks.

Share on WhatsApp

  • 1.75 trillion parameters, versus 175 billion for GPT-3
  • Wu Dao 1.0 announced 11 January 2021; 2.0 on 31 May 2021
  • Trained on 4.9 TB of images and text, incl. 1.2 TB Chinese and 1.2 TB English
  • Wu Dao 2.0 uses a mixture-of-experts design trained with FastMoE
  • Four sub-models: Wen Yuan, Wen Lan, Wen Hui and Wen Su

What it is

Wu Dao, whose name means 'road to awareness', is a multimodal artificial intelligence system built by the Beijing Academy of Artificial Intelligence. Version 1.0 appeared on 11 January 2021 and an upgraded 2.0 followed on 31 May of the same year. It is reported to hold 1.75 trillion parameters, far more than GPT-3's 175 billion. Parameters are the adjustable variables inside a machine learning model. The system handles natural language processing, image recognition, text and image generation, virtual idol animation and protein structure prediction.

How it came about

Work on Wu Dao started in October 2020, a few months after GPT-3 appeared in May 2020. The first release launched large-scale research efforts through four linked models: Wen Yuan, a 2.6-billion-parameter language model for open-domain answering, sentiment analysis and grammar correction; Wen Lan, a 1-billion-parameter image-text model trained on 50 million image pairs for captioning; Wen Hui, an 11.3-billion-parameter generative model aimed at general intelligence problems from a cognitive angle; and Wen Su, a BERT-based model trained on the 100-gigabyte UNIPARC database plus thousands of gene sequences for protein folding.

Inside the model

Wu Dao 2.0 uses a mixture-of-experts design, unlike the dense structure of GPT-3. In such a system only part of the network works on each token, so training costs much less compute than a dense model of the same size, while performance can match models hundreds of times smaller. Training drew on the WuDao Corpora: 3 terabytes of web text, 90 terabytes of graphics including 630 million text-image pairs, and 181 gigabytes of Chinese dialogue covering 1.4 billion dialogue rounds. Overall the model saw 4.9 terabytes of images and text. FastMoE, a Google-published MoE variant, was the training framework.

What it can do

Demonstrations covered natural language processing, image recognition, and generation of both text and images. Wu Dao writes essays, poems and couplets in traditional Chinese, produces alt text for a static picture, and creates near-photorealistic images from written descriptions. It also powered virtual idols with help from Xiaoice, a Microsoft spinoff, and predicted 3D protein structures in the manner of AlphaFold. Commentators called Wu Dao 2.0 the largest language AI system to that date.

How to use it

No public weights, licence or API were described. The model was presented through demonstrations by BAAI, and reporting in June 2021 noted that practical applications and dataset details were still scarce.

Reception and debate

Observers compared Wu Dao with GPT-3 and read it as an effort to compete with the United States in large-scale AI. BAAI's chairman described the goal as building the biggest and most powerful model possible. OpenAI's policy director cited Wu Dao as an instance of 'model diffusion', a term for several groups producing models similar to OpenAI's.

Why it matters

Wu Dao showed that a mixture-of-experts system could reach a trillion parameters and handle text, images and biology at once. Its scale, 1.75 trillion parameters against GPT-3's 175 billion, drew wide attention in 2021 and it was cited as an example of model diffusion, where many groups build models resembling one another.

Questions people ask about Wu Dao

Who made Wu Dao?

It was developed by the Beijing Academy of Artificial Intelligence (BAAI). Development began in October 2020, and the first version was announced on 11 January 2021, with an improved 2.0 version following on 31 May 2021.

How big is Wu Dao compared with GPT-3?

Wu Dao is reported to have 1.75 trillion parameters, while GPT-3 has 175 billion. Parameters are the adjustable variables inside a machine learning model. Wu Dao was trained on 4.9 terabytes of images and text, whereas GPT-3 used 45 terabytes of text.

What can Wu Dao do?

It performs natural language processing and image recognition, writes essays, poems and couplets in traditional Chinese, creates alt text for images, and generates near-photorealistic pictures from descriptions. It has also driven virtual idols with Xiaoice and predicted 3D protein structures.

What is mixture-of-experts?

Wu Dao 2.0 uses a mixture-of-experts design rather than the dense structure of GPT-3. Only part of the network activates for each token, so training needs far less compute than a dense model of equal size, while matching the performance of models hundreds of times smaller.

Can I download or use Wu Dao?

No public weights, licence or API were described. The system was shown through demonstrations by BAAI, and reporting in June 2021 noted that details of training, datasets and practical uses remained scarce.

Good reading and sources

Where this page comes from

The facts on this page come from the Wikipedia article "Wu Dao", written by its contributors and used under the CC BY-SA 4.0 licence, and from its Wikidata record (CC0). EveryLLM rewrote the text in its own words, with AI assistance, and compares every page with the article to make sure no passage is copied. It may be reused under the same CC BY-SA 4.0 licence with a link back.

It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.

The weekly digest

Every Monday: the model of the day, new pages on EveryLLM, and the conferences and anniversaries of the week ahead.