Baidu Unveils Advanced AI Chip Cluster and New Models, Challenges DeepSeek's Dominance

China's Baidu says its Kunlun chip cluster can train DeepSeek-like models

BEIJING/SHANGHAI, April 25 (Reuters) - China's Baidu (9888.HK), opens new tab has successfully "illuminated" a cluster comprising 30,000 of its self-developed, third generation P800 Kunlun chips, which can support the training of DeepSeek-like models, its CEO said on Friday. Robin Li made the announcement at the company's annual developer conference where the Chinese search engine giant is giving updates on its artificial intelligence (AI) efforts. "Illuminated" means switching the cluster on and preparing it for training tasks. The P800 cluster can support the training of DeepSeek-like models with hundreds of billions parameters or a thousand customers fine-tuning models with billions of parameters at the same time, he said. Baidu said Chinese banks and Internet companies had adopted the P800 chips. Li also unveiled Baidu's latest AI model, Ernie 4.5 Turbo, saying it matched the industry's best in several benchmark tests, demonstrating abilities ranging from coding to linguistic comprehension. The company also launched a new reasoning model called Ernie X1 Turbo, and said it would incorporate its AI abilities across its apps from its cloud drive and content platform Baidu Wenku. "There are many (AI) models, but it's apps that rule the world. The application is the king," Li said. "Without apps, models and chips are worthless." The product launch comes at a time of heightened competition in China's AI market, with tech firms shifting focus from foundation model development to the search for applications beyond AI chatbots that can attract and retain users. Baidu was among the first major Chinese companies to invest heavily in AI following the 2022 debut of OpenAI's ChatGPT. However, its Ernie Bot has struggled to gain traction amid fierce competition. Reporting by Che Pan and Brenda Goh; Editing by Himani Sarkar and Kate Mayberry Our Standards: The Thomson Reuters Trust Principles., opens new tab Suggested Topics:Artificial Intelligence Brenda Goh Thomson Reuters Brenda Goh is Reuters' Shanghai bureau chief and oversees coverage of corporates in China. Brenda joined Reuters as a trainee in London in 2010 and has reported stories from over a dozen countries.

Baidu founder highlights 'shrinking' demand for DeepSeek's text-based AI

Baidu's founder has said demand for the type of text-based models developed by generative AI sensation DeepSeek is "shrinking", as his search group seeks to reestablish itself as an artificial intelligence leader in China. In a striking criticism of the limitations of China's AI darling, Robin Li told Baidu's developer conference on Friday that there were constraints to models like DeepSeek's that only generate text-based content. "The market for text models is shrinking," Li said, as he released two new multimodal models -- Ernie 4.5 Turbo and X1 Turbo -- with not just text but also audio, image and video capabilities. He added that DeepSeek's model had a higher propensity for misleading "hallucinations" and was slower and more expensive than other domestic offerings. Li said the competitive landscape for new models was constantly changing, with a stream of "powerful new models that provide more choice". His comments come as Baidu tries to reposition itself as an AI leader after being forced to pivot by dropping its subscription service to its chatbot and making its models freely available as "open source". Baidu faces stiff domestic competition from its peer Alibaba, which has released competitive open-source multimodal models. Baidu presented several use cases for its multimodal models, including an update to its AI avatar platform that enables merchants to create humanlike figures to host livestreams and advertise their products. While Li pointed to DeepSeek's constraints, the internet company has nevertheless embraced its rival's models since the start-up was catapulted into pole position as China's leading large language model player following the release of its R1 reasoning model in January. In recent months, Baidu added DeepSeek to its Qianfan enterprise platform, as well as integrating it into its map and search applications. Charlie Dai, a vice-president at Forrester Research, said Baidu's Friday announcements would "accelerate AI adoption in industries in China, lower barriers for developers and continue to intensify the competition with other leading vendors" such as Alibaba Cloud, Huawei Cloud and Tencent Cloud. Baidu shares in Hong Kong rose more than 4 per cent on the news. DeepSeek is still focused on further developing models, with its engineers working at full steam on the release of next versions -- the R2 and V4 models, the Financial Times reported last month. After the release of ChatGPT in November 2022, Baidu was the first Chinese company to respond to OpenAI's popular chatbot. In March 2023, it announced Erniebot, with the mobile version later rebranded to Wenxinyan. Baidu's chatbot achieved early success in China, but ByteDance's Doubao and DeepSeek's chatbot later overtook it in popularity. This year, Baidu dropped its subscription service after a tepid uptake due to the abundance of free products from rivals. After initially fiercely defending the closed-source model approach, Baidu also started open-sourcing its models, allowing developers greater flexibility to create applications. On Friday, Baidu announced the release of a new AI agent application called Xinxiang, entering an increasingly crowded market that includes Alibaba's Quark app and offerings from start-ups such as Manus AI. Baidu also announced it had built a computing cluster composed of 30,000 Kunlun P800 AI chips from its semiconductor design subsidiary, which it said could support training of several DeepSeek-like models. Li added that developers did not have to worry about a shortage of computing power. The FT last month reported that Samsung had sold Kunlun three years' supply of logic chips, a critical component in manufacturing AI products. Samsung's ability to continue working on Kunlun products has potentially been curbed by new US export controls.

China's Baidu says its Kunlun chip cluster can train DeepSeek-like models

China's Baidu has successfully "illuminated" a cluster comprising 30,000 of its self-developed, third generation P800 Kunlun chips, which can support the training of DeepSeek-like models, its CEO said on Friday. Robin Li made the announcement at the company's annual developer conference where the Chinese search engine giant is giving updates on its artificial intelligence (AI) efforts. "Illuminated" means switching the cluster on and preparing it for training tasks. The P800 cluster can support the training of DeepSeek-like models with hundreds of billions parameters or a thousand customers fine-tuning models with billions of parameters at the same time, he said. Baidu said Chinese banks and Internet companies had adopted the P800 chips. Li also unveiled Baidu's latest AI model, Ernie 4.5 Turbo, saying it matched the industry's best in several benchmark tests, demonstrating abilities ranging from coding to linguistic comprehension. The company also launched a new reasoning model called Ernie X1 Turbo, and said it would incorporate its AI abilities across its apps from its cloud drive and content platform Baidu Wenku. "There are many (AI) models, but it's apps that rule the world. The application is the king," Li said. "Without apps, models and chips are worthless." The product launch comes at a time of heightened competition in China's AI market, with tech firms shifting focus from foundation model development to the search for applications beyond AI chatbots that can attract and retain users. Baidu was among the first major Chinese companies to invest heavily in AI following the 2022 debut of OpenAI's ChatGPT. However, its Ernie Bot has struggled to gain traction amid fierce competition.