Decoding China’s Top Ten Generative AI Models: Established Tech Old Guards Lead the Charge in the GenAI Era

Several hundred large language models (LLMs) have emerged in China, each vying to claim the title of China’s ChatGPT. However, an examination of the top LLMs, whether by valuation, user base, or performance benchmarks, suggests that the generative AI landscape is poised to be dominated by established tech giants, echoing patterns seen in previous waves of innovation.

Leading players such as Tencent and Alibaba possess some of the most robust and capable models. Moreover, they actively invest in new GenAI startups, affording them unparalleled access to top AI talent and expertise in China.

While there are new entrants and unicorns emerging in the generative AI field, many stem from familiar entities like Tsinghua and SenseTime. This underscores the fact that innovation typically springs from existing tech powerhouses. Consequently, generative AI is unlikely to disrupt China’s tech ecosystem but rather solidify the positions of incumbent players.

Below, we delve into the top eight generative AI players in China.

Baidu Ernie Bot (Closed Source)

Founded by: Baidu
Investors: Baidu
Team: Baidu’s TPG (Technology Platform Group) and MEG (Mobile Ecosystem Group) Business units

Valuation: Unknown
Product: Various models & consumer-facing Ernie chatbot
Users: over 200 million
Revenue: Robin Li told analysts during an earnings call in February that "In 2024, the generative artificial intelligence and foundational model business will bring us incremental revenue of tens of billions of RMB, which will also have a positive impact on our total revenue."

Ernie Bot is an artificial intelligence language language model developed by Baidu. Based on Baidu’s Wenxin (Ernie) large-scale model technology, it is a chatbot developed by Baidu after more than ten years of deep cultivation in the field of artificial intelligence. It utilizes multi-modal and cross-lingual deep semantic understanding and generation capabilities, offering users search, document scanning, text-to-image and text-to-video tools. Users can use Ernie Bot for free or pay a subscription to utilize advanced features.

Tencent Hunyuan (Open Source)

Founded by: Tencent
Investors: Tencent
Team: Jiang Jie, Vice President of Tencent Group

Valuation: Unknown
Product: Various models & Hunyuan Assistant
Users: Many Tencent businesses including Tencent Meeting, Tencent Ads, Tencent Games, and Wechat Pay have been connected to the Hunyuan model. Tencent has over one billion users
Revenue: Unknown

Tencent’s Hunyuan model, since its inception, has prioritized practicality and optimization of its own vast business empire. It recently upgraded to a mixture-of-experts model and utilized the DiT architecture, which is employed by OpenAI’s video generation tool Sora. The Hunyuan model has text, text-to-image, text-to-video, and 3D generation capabilities.

Alibaba Qwen (Open Source)

Founded by: Alibaba
Investors: Alibaba
Team: Alibaba’s in-house AI research team

Valuation: Unknown
Product: Various Tongyi Qianwen (Qwen) models
Users: The company said that over 90,000 enterprises adopted their LLM models and over 7 million total downloads
Revenue: Unknown

Alibaba introduced Tongyi Qianwen (Qwen) in April 2023 and have released updated versions. Similar to Baidu and Tencent’s models, these models have text, text-to-speech, and text-to-video generation capabilities. The focus for Alibaba, however, appears to be on attracting enterprise customers to utilize the Qwen models and converting them into loyal users of Alibaba’s cloud computing services. Alibaba Cloud has reported that over 2.2 million corporate users have accessed Qwen-powered AI services, such as its DingTalk platform, which is akin to Microsoft’s Teams.

MiniMax (Closed Source)

Founded by: Yan Junjie, Vice President of SenseTime Technology & Vice Dean of SenseTime Technology Research Institute.
Investors: Alibaba, Tencent, IDG Capital, GL Ventures
Team: Led by a pair of former SenseTime executives. Founders, Yan Junjie and Zhou Yucong, both come from SenseTime. Yan Junjie previously was responsible for building the deep learning toolchain and underlying algorithms and Zhou Yucong was formerly in charge of the algorithm team at SenseTime.

Valuation: ~ US$2.5 billion
Product: Various text and speech models, as well as consumer-facing tools Wanjuan, Xingye, and Hailuowenwen.
Users: Glow reportedly had 5 million users, and Xingye was reported to have users in the millions.
Revenue: Unclear

MiniMax has both a to-consumer and to-business strategy. It builds underlying foundational models, having released the text large language model series abab and the speech model abab-speech-01. It unveiled consumer-facing products including the AI chatbot Glow, AI assistant Wanjuan, the AI social app Xingye, and the AI voice conversation assistant Hailuowenwen.

ByteDance Skylark

Founded by: ByteDance
Investors: ByteDance
Team: ByteDance’s in-house AI team

Valuation: Unknown
Product: Various models and numerous consumer-facing apps including Doubao, Cici, Coze, KouZi, and others.
Users: Undisclosed
Revenue: Unknown

ByteDance’s approach to generative AI appears to resemble casting a wide net, experimenting with numerous applications in the hopes of finding the next viral sensation akin to TikTok. The company has introduced at least seven consumer-facing generative AI apps, such as Doubao and Cici, with many of them available in both Chinese and international versions. Despite this extensive portfolio, none of these AI applications have yet achieved substantial growth.

Conversely, ByteDance is making significant strides in AI-driven image and video tools, leveraging its proprietary models developed within its Volcengine unit. This emphasis on AI technologies is rooted in ByteDance’s expertise in short-form videos and content-rich platforms, as it endeavors to maintain its dominance in the evolving landscape of generative AI.

SenseTime’s SenseNova (Open Source & Closed Source)

Founded by: SenseTime
Investors: SenseTime
Team: SenseTime in-house research team

Valuation: Unknown
Product: Various SenseNova LLMs
Users: Unknown 
Revenue: The SenseNova LLMs-driven business has experienced an explosive growth of 200%, with revenue reaching 1.2 billion yuan in 2023, contributing 35% of the group’s total revenue, SenseTime reported in March.

SenseTime’s strategy focuses on large-scale models and providing AI computing power. It has released SenseNova large language models offering capabilities such as natural language processing, content generation, automated data annotation, custom model training, AI-generated images and graphics, 2D/3D digital human generation, and large scene/small object generation.

Baichuan (Open Source & Closed Source)

Founded by: Wang Xiaochuan, founder of search engine Sogou
Investors: Alibaba, Tencent, Xiaomi,
Team: Led by former CEO of Sogou, Wang Xiaochuan, Baichuan’s core team consists of AI talents from Sogou, Google, Tencent, Baidu, and other big tech companies.

Valuation: ~US$1.8 billion
Product: Various models
Users: Unknown
Revenue: Unknown

Baichuan has released four open-source large language models, namely Baichuan-7B/13B and Baichuan2-7B/13B, which are free for commercial use. Additionally, it has launched two proprietary closed source models, Baichuan-53B and Baichuan2-53B. Baichuan has opened the API interface for Baichuan2-53B, signaling it has officially entering the B2B field and commencing the commercialization process. It recently released the character-building model Baichuan-NPC, which is optimized for "character knowledge" and "dialogue capabilities".

01.AI (Open Source & Closed Source)

Founded by: Dr. Kai-Fu Lee, venture capitalist and formerly with Google and Microsoft
Investors: Alibaba Cloud, Sinovation Ventures (also founded by Dr. Kai-Fu Lee)
Team: Led by Dr. Lee, the team includes talents from Alibaba, Baidu, Didi and Google.

Valuation: ~US$1 billion
Product: various Yi models
Users: Unknown
Revenue: Close to profitability, according to Dr. Lee

01.AI has released both open source and closed source models, including Yi-34B-Chat, Yi-6B-Chat, Yi-34B, Yi-34B, Yi-9B, and Yi-6B. It released closed-source Yi-Large model with 100 billion parameters. 01.AI is mainly focused on global deployment and the integration of models and applications. Recently, it launched its consumer-facing AI service called Wanzhi, and has started training the next generation Yi-XLarge MoE model.

Dr. Lee said that the company is closed to profitability, as the company competes with lower price. Its paid API costs only US$2.5 for inputting 1 million tokens and US$12 for outputting 1 million tokens, lower than competitors such as OpenAI.

Zhipu AI (Open Source)

Founded by: Knowledge Engineering Group at Tsinghua University
Investors: Social Security Fund, Meituan, Ant Gorup, Alibaba, Tencent, Xiaomi, and others
Team: CEO Zhang Peng graduated from the Department of Computer Science and Technology at Tsinghua University, with his doctoral research focused on knowledge graphs. Chairman Liu Debing previously served as Deputy Director of the Technology Big Data Research Center at Tsinghua Data Science Research Institute. President Wang Shaolan holds a Ph.D. from Tsinghua University.

Valuation: US$1.4-US$2.8 billion
Product: AI assistant, code model CodeGeeX, multimodal CogVLM, and the text-to-image model CogView
Users: Over 1,000 enterprise adoptions of Zhipu AI’s models
Revenue: Signed contracts worth US$14 million

Zhipu AI is dedicated to building a new generation of cognitive intelligent large models in China. The company has collaborated to develop the GLM-130B, a bilingual Chinese-English large-scale pre-training model, and launched the ChatGLM dialogue model based on this, along with the open-source single-card version model ChatGLM-6B. At the same time, the team has also developed the AIGC model and product matrix, including the AI assistant, the high-efficiency code model CodeGeeX, the multimodal model CogVLM, and the text-to-image model CogView. The company adopts the concept of Model as a Service (MaaS) and has launched the large-scale model MaaS open platform.

Moonshot AI (Closed Source)

Founded by: Zhilin Yang, Tsinghua University graduate and a Ph.D. in computer science from Carnegie Mellon University. He is the first author of two papers, Transformer-XL and XLNet
Investors: Alibaba, Sequoia Capital, Meituan, Xiaohongshu
Team: Zhilin Yang and two co-founders, who are alumni with Yang.

Valuation: US$2.5 billion
Product: Consumer-facing chatbot KimiChat
Users: 890,000 unique visits
Revenue: Unknown

Moonshot AI has released one consumer product, a chatbot called KimiChat. The chatbot was marketed as the most capable of processing long text, with the company claiming itself as the first intelligent assistant product capable of supporting input of up to 200,000 Chinese characters. Users say that the chatbot can read long texts, provide summaries, and engage in multi-round question and answer sessions. Moonshot AI is also currently developing a multimodal model and is expected to launch it later this year.


Related News: