Der ehemalige Technologieleiter von ByteDance gründet ein Start-up und möchte eine Unternehmens-Coding-Agent-Plattform aufbauen. Bereits wurden mehrere Millionen Yuan an Finanzierungen erhalten | Exklusives Interview von 36Kr
Text | Deng Yongyi
Editor | Su Jianxun
In just half a year, Yang Ping witnessed how crazy the AI Coding track was.
In 2024, the Vibe Coding track was booming. Cursor's ARR soared from $1 million in 2023 to $65 million in November 2024, and its valuation more than sextupled in just four months.
At that time, Yang Ping had just launched the first Coding product, MarsCode, into the market at ByteDance. As the person in charge of technology R & D, since 2021, she had led a team of hundreds of people at ByteDance to explore the application of AI in the software field. The products she created had served tens of thousands of R & D personnel within ByteDance, helping the company save hundreds of millions in R & D budgets.
"Even then, we had already discovered that AI had great potential for application in the Coding field, and users were more concerned about whether they could directly understand and effectively use it." She told Intelligent Emergence.
However, when the track changed drastically, Yang Ping felt a strong sense of urgency.
The changes were reflected in two dimensions: Firstly, these consumer - oriented Coding products not only received high valuations from capital but also saw a sharp increase in revenue, proving the real market demand for AI programming tools.
Secondly, the market landscape was being reshaped. Whether in China or overseas, in the consumer - oriented Coding product market, the pioneers had already staked their claims. However, more and more manufacturers were turning to the business - to - business (B2B) market, and the enterprise - level demand was starting to explode.
These changes made her realize that the time for entrepreneurship could not be delayed any longer. In July 2025, Yang Ping and two co - founders officially established a new company, "Ciyuan Infinite", hoping to provide AI Coding Agent services for B2B enterprises.
Recently, Ciyuan Infinite officially completed an angel - round financing of tens of millions of yuan. The investor was a CVC of a software industry. This round of financing was advised by Navigation Capital as the long - term financial advisor.
Ciyuan Infinite has also recruited senior talents from prestigious institutions such as the Yao Class of Tsinghua University and companies like ByteDance. The CTO, Wang Wei, is an alumnus of the Yao Class of Tsinghua University and was formerly a technology partner of a well - known domestic embodied robot company and a large - model startup. The person in charge of commercialization, Li Ying, has more than a decade of experience in the implementation of AI industries and has led the implementation of projects worth hundreds of millions of yuan in the B2B field of large models.
Although general AI programming tools on the market can generate code quickly and write many small applications rapidly, they often struggle when faced with the complex legacy systems (Legacy Code), technology stacks, and strict business specifications of enterprises.
For example, in a typical financial scenario: The China Banking and Insurance Regulatory Commission stipulates that the account - opening process must include specific compliance steps, but general AI tools may design several steps on their own based on public data. No matter how fast the code is generated, in the eyes of enterprises, it is useless and discarded code.
To truly integrate Vibe Coding into enterprise - level production scenarios, Silicon Valley is reviving a classic concept - FDE (Forward Deployed Engineer).
This model was proposed by Palantir in 2010 and has currently sparked a recruitment boom in the AI industry. Enterprises including OpenAI and Anthropic have announced that they will expand their application AI teams, including FDEs, by several times.
Ciyuan Infinite hopes to use AI to create a similar Agent service, enabling high - level engineers to directly enter the customer's site. They are both technical experts and business consultants, helping enterprises apply cutting - edge AI technology to the production environment.
InfCode
At the beginning of December, the first version of Ciyuan's core product, InfCode, was officially launched. It is in the form of a plugin + an enterprise - level AI Coding platform. InfCode helps enterprises complete tasks such as code governance, completion, review, and task planning, equivalent to a mid - level R & D engineer.
InfCode solves this problem through a two - layer mechanism: The first layer is standardized docking. Through the built - in MCP Server connector, it can quickly integrate common office systems such as Feishu and enterprise OA, enabling AI to query enterprise - internal documents and specifications in real - time.
The second layer is personalized adaptation. For each enterprise's unique microservice architecture, internal processes, and legacy systems, it provides open interfaces for the enterprise's IT team to perform lightweight adaptation.
In actual POC verification, in the current customer cases that Ciyuan is cooperating with, the R & D efficiency has increased by nearly 40%, the usability rate of AI - generated code has reached over 88%, and the quality has reached the level of an intermediate programmer.
More importantly, Ciyuan Infinite does not focus on the accuracy of AI in the intermediate process but directly measures the value of project delivery, such as the improvement of human efficiency and project time.
Currently, on the globally authoritative agent evaluation benchmark, SWE - Bench Verified, InfCode has refreshed the world's best record (SOTA) with a score of 79.4%, exceeding the scores of top models such as GPT - 5 and Claude on the public leaderboard by about 65%.
In China in 2025, entering the B2B market is still a challenging decision. In the previous AI 1.0 era, the development of many AI companies showed that if a company only has single - point technical capabilities, it is easy to get trapped in endless customization and large - scale projects.
However, Yang Ping believes that AI Coding may pave a different path for B2B implementation.
Behind this, one reason is the rapid development of base models. In 2025, base - model manufacturers have spared no expense in building Agent capabilities. The newly launched Gemini 3 flash is already several times more powerful than Gemini 2.5 pro a few months ago, bringing immediate business value.
Secondly, consumer - oriented AI products are rapidly spreading. For example, vertical - category Coding products like Cursor and Replit, as well as general AI assistants like ChatGPT, have far exceeded expectations in terms of the enterprise R & D community's acceptance and willingness to pay for AI.
Yang Ping said that ultimately, the business model of Agent will move towards a result - oriented approach (RaaS, Result as a Service), enabling enterprise managers to clearly calculate the return on investment. "The popularization of consumer - oriented products has already completed the market education of enterprise R & D personnel. Now the key is to solve the problem of business understanding in the last mile." Yang Ping said.
Upon the completion of the first - round financing, several co - founders of Ciyuan Infinite also talked with Intelligent Emergence about the pain points and opportunities in B2B AI Coding.
The above is the original text of the interview, edited and organized by Intelligent Emergence.
Enterprise - level scenarios and Vibe Coding are largely incompatible
Intelligent Emergence: Let's start from the origin of your entrepreneurship. Initially, you were working on Coding products at ByteDance, and then you started your own business. What were your considerations behind this whole process?
Yang Ping: From 2018 to 2024, I was mainly responsible for work related to intelligentization and software engineering at ByteDance. Before the popularity of large language models, we had already been researching some deep - learning and generative - learning models, but at that time, we were mainly exploring in the academic field.
Since 2021, I and the entire team have continuously observed and carried out relevant applications of reinforcement learning, including the application of deep learning in code scenarios.
ByteDance's internal AI Coding product started in mid - 2022, and we basically experienced the entire process from scratch. Later, I chose to leave in August 2024. After a period of study, I founded Ciyuan Infinite in July this year.
Intelligent Emergence: In August 2024, MarsCode had just been launched. Did you hesitate to leave the product that you had nurtured from the start?
Yang Ping: The most crucial factor is the numerous opportunities brought by the wave of large models. One should make choices at the peak rather than in a downturn. Otherwise, what reason do you have to convince others to invest real money in you?
An entrepreneur should have the mentality to go all - out when things are going well. At this time, one has both experience and confidence.
Before we established our company, star companies like Cursor and Replit started to see a sharp increase in their market performance and valuations in the second half of 2023.
This really touched me. I realized that in a large company, the competition might not be as fast - paced as in the outside world. The timing of entrepreneurship is crucial.
Intelligent Emergence: Previously, the AI Coding products you developed were directly targeted at developers, but your new company, Ciyuan Infinite, has chosen the enterprise - level market. Why did you decide to change the direction?
Yang Ping: The difference between B2B and B2C in developer tools is not as significant as that in traditional software products. When we emphasize enterprise - level, we mainly refer to the problem domain we want to address. Enterprise - level scenarios require large - scale Agent capabilities to solve problems throughout the entire software - delivery process.
For example, many consumer - oriented Vibe Coding products claim to be able to generate a demo project of tens of thousands of lines of code with just one sentence. However, it's hard to believe that Vibe Coding can generate a software project of tens of millions of lines that can increase revenue or support core business.
We actually focus more on serious enterprise - level programming scenarios and aim to tackle more complex and large - scale problems.
Intelligent Emergence: Can you summarize in one sentence what Ciyuan Infinite's products can do?
Yang Ping: We are an AI Coding Agent platform focused on enterprise - level scenarios.
From a fundamental perspective, we offer not only a tool but also platform - level Agent capabilities. These Agents will be deeply integrated into every aspect of enterprise - level software development, such as enterprise - level Coding, Debugging, etc., and deeply integrate AI capabilities into the IDEs (development environments) commonly used by developers.
Intelligent Emergence: In enterprise - level development scenarios, why can't Vibe Coding products meet the requirements? What are the user pain points?
Yang Ping: Enterprise developers often find it very troublesome to frequently switch between multiple platforms such as requirements, development, testing, and operation and maintenance. There are huge data and process barriers in between. For example, in key industries such as finance and medicine, every code change must be accountable for the business.
In serious programming scenarios, the results are relatively certain, and service stability is required. It is necessary to guide users through the entire Coding process according to specific context.
So, in essence, this is incompatible with Vibe Coding. Vibe Coding products often rely more on the capabilities of base models, and the engineering layer is relatively thin. It is not just a reflection of personal will. One has to understand business requirements, write business code according to business specifications, and carry out projects under certain constraints.
Intelligent Emergence: Can you give an example to illustrate the actual difficulties encountered by enterprise customers and how you solve them?
Wang Wei: Enterprise development often involves the collaboration of dozens or even hundreds of people. Therefore, the product must be able to naturally integrate with the enterprise's existing R & D processes.
For example, when developing the front - end pages of an e - commerce system, if only Vibe Coding is used, the pages written may not match the logic of the back - end system. However, our AI Coding Agent can connect to multiple back - end projects, refer to the corresponding code interfaces and documents, and ensure that the front - and back - end code logic is consistent and can actually run.
The limited effectiveness of Vibe Coding products in B2B scenarios is precisely because they are not designed for such scenarios.
Take the account risk - control logic in finance as an example. The China Banking and Insurance Regulatory Commission has clear policy - guidance documents requiring that the account - opening process must include several inspection steps. If a simple Vibe Coding product is used for generation, it may "intelligently" design a very complex inspection process, but the final result will not meet the basic process requirements at all.
Our Coding Agent aims to solve scenarios that are complex, involve a large amount of code, and have many associations. For example, in an enterprise's "existing projects", the code repositories may have hundreds of thousands or even millions of lines of code, and the context space of the model is definitely insufficient.
This requires optimization at the Agent level. We have designed two mechanisms for our Agents. Internally, we mainly focus on optimizing the context under the limited context window, including dynamic context compression and loading/unloading mechanisms, to address the inherent limitations of the model. Externally, we mainly introduce more connections, dynamically bringing various types of information (including documents, external dependencies, databases, etc.) in the enterprise R & D process into the Agent environment through open protocols such as MCP. This combination of internal and external optimization has achieved good results.
Intelligent Emergence: Can you give a specific case to illustrate how your product helps users solve problems?
Wang Wei: We currently have a typical cooperation case with a listed financial company. The implementation is divided into two steps:
The first step is to provide a standardized product, focusing on context engineering to ensure that the design of the intelligent agent itself can avoid the problem of insufficient context length.
The second step is to solve the problem of "where the information comes from".
This information docking is divided into two categories. One is standardized. Through the built - in MCP Server connector, it can quickly integrate environments such as git repositories, artifact - management systems, and document systems, enabling the Agent to accurately read external information.
The other part is non - standard personalized systems. We provide open interfaces and adapters, and the customer's team has quickly integrated their internal OA system.
After these two steps of implementation, the final result is both fast and good. In the customer's actual test, the product's ability is basically on par with that of an intermediate developer, and ultimately helped the customer increase human efficiency by nearly 40%.
Intelligent Emergence: How is this 40% calculated? How is a problem considered solved?
Wang Wei: We only look at the final result, that is, we directly compare the human - resource input, such as the R & D cycle (person - days).
These indicators must be directly related to business goals. During the entire development process, we do not break down the different types of situations in the middle, nor do we care how much human resources are needed to fix or adjust the code after AI generation. We only look at how much human - resource cost is saved when the team completes the same - value tasks.
From the perspective of enterprise managers, they are most concerned about the final delivered value, rather than other process indicators.
Intelligent Emergence: So, what is your charging model?
Li Ying: Currently, we are in the process of signing agreements and exploring cooperation with several leading ISVs (Independent Software Vendors) and large - scale end - customers.
For tool - type products, we will charge license fees and Agent subscription fees. For platform - type products, in addition to charging some standard fees, we are also considering a profit - sharing model.
From plugins to IDEs, the product form is not the most core issue
Intelligent Emergence: You just mentioned that Ciyuan Infinite aims to solve "serious programming scenarios". What are the main differences between the problems it solves and those solved by Vibe Coding?
Li Ying: The Coding products in the consumer market on the market actually address the delivery capabilities of lightweight software. However, when facing a complex enterprise - level business scenario, they often struggle to perform better.
And this is precisely the huge market space in our eyes. It may seem like difficult work, but it has become