The landscape of China’s top domestic AI startups, once defined by the “Five Tigers” of large language models (LLMs), is undergoing a significant shift.
A low-profile but rapidly advancing company is now stepping into the spotlight with greater momentum.
StepFun, an AI LLM company headquartered in Shanghai and founded by Jiang Daxin, former Global Vice President at Microsoft, is gaining increasing recognition following the debut of its self-developed models and application products, building a growing “snowball” effect.
This site has learned that in the latest financing market, StepFun has become a favorite, currently conducting a new funding round with a valuation of $2 billion.
Interestingly, Alibaba once again appears in rumors regarding potential layout or investment.
In previous investor lists for StepFun, one can see the presence of top-tier venture capital firms such as Qiming Venture Partners and Matrix Partners China.
The term “Five Tigers” was previously an industry shorthand for the leading unicorn startups in China’s domestic LLM sector. These five companies are Zhipu AI, Moonshot AI, 01.AI, Baichuan Intelligence, and MiniMax.
Now, with StepFun’s strong rise, this collective title may need to be updated.
In the future landscape of Chinese foundational model startups, attention may no longer focus solely on the “Five Tigers,” but rather on a new group dubbed the “Six Little Giants.”
(If you have better naming suggestions, please leave a comment below.)
Dual Focus on Technology and Products
StepFun made its public debut relatively late.
Founded in 2023, it did not officially reveal itself to the public until late March this year.
However, over the past three months, this domestic LLM startup has been actively showcasing its capabilities in both technology and product development.
It has also outlined its long-term roadmap toward Artificial General Intelligence (AGI):
Single-modal → Multi-modal → Unified multi-modal understanding and generation → World models → AGI.
Step Series General-Purpose Large Models
Accompanying StepFun’s debut were three large models developed by the company:
- Step-1: A 100-billion-parameter language model
- Step-1V: A 100-billion-parameter multi-modal model
- Step-2: A trillion-parameter MoE (Mixture of Experts) language model
Among these, the 100-billion-parameter Step-1 was successfully trained in just two months, with comprehensive performance surpassing GPT-3.5.
Two months later, in November last year, the 100-billion-parameter multi-modal model Step-1V was completed. The Step-1V model can accurately describe and understand text, data, charts, and other information within images, and perform tasks such as content creation, logical reasoning, and data analysis based on image inputs.
Additionally, it is capable of understanding video content.
In March this year, StepFun became the first domestic startup LLM company to publicly unveil its trillion-parameter model, Step-2, which utilizes MoE technology.
Two ToC Products
To date, StepFun has publicly released two consumer-facing (ToC) products, both open to all users.
One is called Yuewen (Leap Ask), a chat-based application positioned as a personal efficiency assistant.
It supports uploading external documents, features internet connectivity and information retrieval capabilities, possesses multi-modal abilities, and can summarize charts in its outputs.
It is currently available for free use on web browsers and mobile apps.
The other product is called Maopaoya (Bubble Duck), an AI open-world platform offering a vast array of intelligent agents, primarily focused on leisure and entertainment.
Interested readers are encouraged to experience it firsthand; further details are omitted here.
Additionally, the LLM-native application “The Player: I Broke the Large Model” (Touhao Wanjiia), which this site recently shared with readers, is also powered by StepFun’s large model technology.
△ On the Ten Thousand Ways to Draw a Penguin
The Founding Team from Microsoft
Finally, let’s briefly review the key leaders of StepFun.
CEO Jiang Daxin also serves as the team’s lead algorithm expert.
After earning his Ph.D. in Computer Science from the University at Buffalo (SUNY), Jiang served as an Assistant Professor at Nanyang Technological University and later as a researcher at Microsoft Research Asia (MSRA).
Starting in 2011, he joined Microsoft Asia Internet Technology Engineering Institute (STCA). In March 2023, he was promoted to Vice President of Microsoft, concurrently serving as General Manager of the WebXT S+D group (Web Experience, Search, and Distribution) at the Asian Software Technology Center.
But life offers stages everywhere.
Vice President Jiang Daxin quickly developed the idea to embark on LLM entrepreneurship and acted on it immediately.
Jiao Binxing, Head of Data at StepFun, left Microsoft alongside Jiang Daxin to start this venture.
Jiao is a joint Ph.D. candidate from the University of Science and Technology of China (USTC) and MSRA. After graduation, he joined Microsoft, eventually rising to become the former head of Bing’s core search team.
His daily work primarily involved using data mining and NLP algorithms to optimize indexing and search quality.
Public records indicate that he also developed automatic mining algorithms for high-quality global websites, which were applied to indexing and ranking.
Zhu Yibo is another key member of the core team, currently serving as Head of Systems at StepFun.
After graduating from the University of California, Santa Barbara (UCSB), he joined Microsoft before moving on to ByteDance and Google, gaining extensive practical experience in large-scale systems and 10,000-GPU clusters.
It can be said that the backgrounds of these three core members are quite similar: Phase one involved relevant research during their doctoral studies, followed by Phase two where they worked at major tech companies on the front lines of AI.
More notably, they all emerged from Microsoft, forming a “Microsoft-affiliated” entrepreneurial team.
Not long ago, it was also reported that Microsoft Senior Principal Researcher Duan Nan left MSRA to join StepFun.
With more talented individuals joining, will this intensify the competition among the “Six Little Giants”?
In other news, OpenAI began sending emails to many developers starting last night, stating:
Additional measures will be taken starting July 9 to stop API usage from countries and regions not on its supported list.
Fortunately, we have domestic large models to rely on…
The “Six Little Giants” of LLMs must keep pushing forward!
Comments
Sign in to join the discussion and leave a comment.
Sign in with Google