Open-Source Models and New Funding: How China's AI Start-Ups Are Responding to DeepSeek

The logo of DeepSeek was showcased during the Global Developer Conference held in Shanghai, China, on February 21. The emergence of DeepSeek has caused significant ripple effects throughout China's artificial intelligence (AI) landscape as its affordable models challenge traditional industry perspectives. This disruption has pushed numerous startups to explore open-source model development and pursue fresh capital infusion.

DeepSeek, a company that evolved from a hedge fund led by Liang Wenfeng, has made a considerable impact not only in China but also on platforms like Wall Street and Silicon Valley. Based in Hangzhou, the startup has captured considerable media attention and outshined local competitors like Moonshot AI and MiniMax.

In July 2024, the launch of DeepSeek's V2 model sparked an intensifying price competition within China's AI sector. Liang described the company’s unexpected role as a "catfish" in revolutionizing the industry, noting that their intention was not to disrupt the existing order. However, the newer releases of V3 in December and R1 in January raised challenging questions about survival for many businesses in China's saturated AI model market.

The ripple effect from DeepSeek has seemingly provided China with an advantage in the realm of advanced AI capabilities, ultimately democratizing access to this technology so that more companies can foster innovation.

Grace Shao, an AI analyst and founder of the industry newsletter AI Proem, explained that prior to the introduction of DeepSeek’s R1, many Chinese AI startups were pivoting toward consumer applications, driven by monetization strategies from the mobile internet era. In contrast, the U.S. AI landscape has largely embraced AI to enhance enterprise functionality and boost white-collar productivity.

Shao attributed the differing methodologies to the structural economic contrasts between Chinese and U.S. markets, emphasizing that having robust model capabilities is essential for building a strong AI industry.

Consequently, other AI developers in China are striving to catch up with DeepSeek by seeking significant advancements in AI technology.

On Monday, Zhipu AI, a startup from Beijing that received support from Tsinghua University back in 2019, announced it had raised 1 billion yuan (approximately $140 million). This funding included contributions from Hangzhou's municipal government. Zhipu AI, which is also backed by notable investors like Qiming Venture Partners and Hillhouse Capital, is embracing the open-source movement by releasing its AI models and agents for developers to utilize. Recently, they unveiled an open-source text-to-image model named CogView-4 that can generate Chinese characters.

The open-source model enables broader access to a software’s source code, allowing developers to modify, share, or enhance capabilities. The open-source philosophy has significantly contributed to the development of China’s competitive tech industry in recent years.

"Entrepreneurs from the 1980s and 1990s aspire to show the world that Chinese firms can innovate rather than merely imitate," Shao noted, explaining the current trend toward open-source developments in AI.

She remarked, "It is more gratifying to have your work cited or utilized by developers and businesses beyond China's borders than just generating revenue from one project."

Shanghai-based startup Stepfun, founded in 2023 by former Microsoft Research Asia chief scientist Jiang Daxin, launched two open-source multimodal models last month: Step-Video-T2V, which converts text into videos, and Step-Audio for voice interactions. The startup has plans to release an image-to-video model soon.

Stepfun's funding backers include the municipal government's Capital Investment of Shanghai, the internet giant Tencent Holdings, Qiming Venture Partners, and 5Y Capital, as stated in their latest funding announcement at the end of 2024.

The AI startup MiniMax, recognized for its personalized AI applications named Talkie and the domestic counterpart Xingye, began their foray into open-source model development shortly after the launch of DeepSeek's V3 in January. The company rolled out its large language model (LLM) MiniMax-Text-01 and a corresponding multimodal model known as MiniMax-VL-01—a technology often utilized in generative AI services like ChatGPT and Baidu's Ernie Bot.

In an interview, founder Yan Junjie remarked, "If I could reset, I would have advised an open-source approach from the start."

Meanwhile, Moonshot AI, known for the Kimi chatbot, released a multimodal reasoning model termed K1.5 in January, coinciding with the unveiling of DeepSeek's R1. Recently, Moonshot AI has launched its own open-source architecture and optimization innovations.

Baichuan AI, established by ex-Sogou CEO Wang Xiaochuan, has redirected its focus toward the healthcare sector. Recently, the company made the decision to let go of its financial services team, confirming the restructuring to the South China Morning Post while asserting the intention to concentrate its resources on core medical ventures.

01.AI, co-founded by former Google China head Lee Kai-fu, has transitioned from developing extensive AI models to concentrating on applications tailored for specific industries. Earlier this year, they collaborated with Alibaba Group Holdings' cloud service department to establish a joint lab focused on industrial-scale AI models, with some of 01.AI's team members transitioning to Alibaba Cloud as part of this partnership.

In conclusion, as various startups respond to DeepSeek's impact, a wave of innovation is taking place across the Chinese AI sector as they explore open-source models, attract new investment, and aim to demonstrate their unique capabilities on both domestic and global stages.

AI, China, DeepSeek