TikTok parent ByteDance has launched an updated version of Doubao, China’s most popular consumer-facing artificial intelligence (AI) app, as the tech giant accelerates AI development despite US export restrictions on advanced chips.
The Beijing-based company introduced its closed-source multimodal model Doubao 1.5 Pro on Wednesday, emphasising a “resource-efficient” training approach that it said does not sacrifice performance.
“The model adopted an integrated train-inference design from the pre-training phase to balance between the best performance and most optimal inferencing cost,” ByteDance said in a statement, adding that it has designed a server cluster with flexible support for low-end chips to bring down the AI infrastructure costs.
Do you have questions about the biggest topics and trends from around the world? Get the answers with SCMP Knowledge, our new platform of curated content with explainers, FAQs, analyses and infographics brought to you by our award-winning team.
China’s Big Tech firms are striving to catch up with their US counterparts while facing budget constraints and limited access to advanced chips. This has pushed them to innovate in AI model efficiency, refining their products within the country’s closed market.
ByteDance’s ChatGPT-like Doubao bot has emerged as the most popular AI app for Chinese consumers. Photo: Simon Song alt=ByteDance’s ChatGPT-like Doubao bot has emerged as the most popular AI app for Chinese consumers. Photo: Simon Song>
Benchmark tests have shown that Doubao 1.5 Pro excels in half of 14 evaluations that assessed the model’s language understanding, maths and coding skills, domain knowledge, visual understanding and reasoning abilities.
In some areas it outperformed industry-leading AI systems from Microsoft-backed OpenAI, Google, and Amazon.com-backed Anthropic. It also bested domestic rivals in some tests, including systems from recent start-up darling DeepSeek and cloud computing giant Alibaba Group Holding, owner of the Post.
ByteDance’s resource-efficient AI development strategy follows the success of Hangzhou-based DeepSeek, which stunned the global tech community with its V3 model unveiled late last year. DeepSeek’s open-source V3 was trained using significantly fewer resources than Meta Platforms‘ Llama 3.1, while delivering comparable performance.
DeepSeek disclosed in the model’s technical report that it was trained using Nvidia’s H800 chips at a cost of less than US$6 million. By comparison, Anthropic CEO Dario Amodei estimated that OpenAI spent US$100 million on training GPT-4o.
ByteDance did not disclose the cost or chips used for training its latest model, and did not immediately respond to a request for comment on Wednesday.
The ability to advance AI models while managing resource constraints is playing a more important role in the industry amid an escalating US-China tech rivalry.
Despite US restrictions, Chinese companies continue to make strides in AI development. The Doubao 1.5 Pro release came days after DeepSeek and Beijing-based Moonshot AI launched new reasoning models this week.
This article originally appeared in the South China Morning Post (SCMP), the most authoritative voice reporting on China and Asia for more than a century. For more SCMP stories, please explore the SCMP app or visit the SCMP’s Facebook and Twitter pages. Copyright © 2025 South China Morning Post Publishers Ltd. All rights reserved.
Copyright (c) 2025. South China Morning Post Publishers Ltd. All rights reserved.
