Meituan opens AI code tools to achieve full-stack capabilities with zero code, and the project leader reveals the architecture details

No one expected that such a practical AI code generation tool would come from Meituan.

Last week, some media exposed Meituan’s AI zero-code tool NoCode, which is a tool that can quickly generate applications through natural language and conversation without requiring programming background and experience.

As the name suggests, NoCode can help many people create personal efficiency tools, product prototypes, interactive pages, etc. in a "zero-code" way. It can not only generate code, but also preview in real time, make partial modifications and deploy with one click, which greatly reduces the threshold of development and can help more people unleash their creativity.

Moreover, NoCode is completely free and you can log in by scanning the QR code using the Meituan App or WeChat.

• Product link: https://nocode.cn/

NoCode is the latest practice of Meituan's open AI ecosystem. It aims to help small and medium-sized businesses achieve IT and digital upgrades by opening up its own accumulated AI technical capabilities for free, while allowing more users to experience the efficiency improvement and innovative fun brought by AI technology. Within the company, people have used it to build a large number of different types of applications, from website pages to efficiency tools, data analysis to simple games.

Although it has not yet been officially released, we have already seen some use cases of products built using NoCode on social networks.

For example, the Gobang game generated by NoCode has visual feedback when the pieces are placed, and can be played by humans and machines:

Project link: https://nocode.host/6281d/wqtkm8blcfm3ovhk

A website for tracking hot food, with detailed distinctions between categories, store details, and other pages, including store coverage heat maps:

Project link: https://nocode.host/3a6ad/m3dbk1k0j4ksr99s

Merchant digital operation system, including operation data, order management, inventory and analysis, etc.:

Project link: https://nocode.host/6281d/1o5c1i2i9bdacm06

More and more people are sharing their experience. In short, NoCode has the following features:

• Natural language programming: Use natural language to describe ideas, and NoCode will automatically interpret and convert them into complete functions, generating usable capabilities without programming experience. After the user enters the command, NoCode will expand and optimize the command, and can also perform one-click optimization.

• Real-time preview effect: NoCode can instantly render and present results based on the conversation content, and you can view the actual effect after each conversation in real time.

• Local positioning modification: Use the Visual Edit function to make local modifications and improvements to the positioning content; it also supports comparison and rollback between versions to ensure that every step is "traceable".

• One-click deployment and sharing: After the application is built, the code will be automatically uploaded to the warehouse, and can be deployed immediately and shared with others. Click to directly access the link to view the work. At the same time, a copy button will be displayed at the back. Users can also choose to copy the link with one click and share it with others for viewing.

NoCode was developed by Meituan's R&D Quality and Efficiency Team, which is part of Meituan's Basic R&D Platform. Before NoCode was officially launched, Cheng Datong, the person in charge of engineering efficiency of Meituan's Basic R&D Platform and a member of Meituan's Technical Committee, and Yu Chao, the person in charge of intelligent development tools of Meituan's Basic R&D Platform, accepted our interview. They gave us a detailed introduction to NoCode's product form, technical architecture, and application capabilities.

In the eyes of NoCode's developers, this tool is like a full-stack engineer, which can improve productivity in different workflows. From a technical architecture perspective, NoCode is based on multiple AI models, and has been specially trained and tuned for multiple businesses. In Meituan's internal applications, NoCode has demonstrated strong practicality. Perhaps it won't be long before our programming paradigm will undergo a radical change because of these AI tools.

This is the first time we can learn about the technical details of this mysterious product from Meituan. Let’s see what they say.

Some texts have been rewritten without changing the original meaning.

New tool releases

1. Please briefly introduce Meituan’s upcoming AI zero-code tool NoCode.

Cheng Datong: To some extent, you can think of NoCode as a full-stack AI engineer. It has a complete environment and a sandbox that can run programs directly.

Even if you don’t have a technical background, you can still generate gadgets, web front-ends, product prototypes, applets, mini-games, and more. You only need to present your requirements to AI, and it will do the work for you without bothering the development team or considering installation and deployment. Within Meituan, many students who have no programming experience have built applications with actual productivity through multiple rounds of dialogue with AI. Although they are not as complex as those built by professional developers, they are already fully usable.

I think it means creativity: if you have an idea in your professional field, NoCode may be able to help, but you need some patience. Within Meituan, some excellent practice cases made with NoCode will have more than a hundred rounds of dialogue with AI.

2. We know that Meituan has been exploring the application of new technologies in retail scenarios for a long time. The product proposed this time is oriented towards code development, which can be said to have opened up a new direction. What are the considerations for opening NoCode?

Cheng Datong: We think we have reached this stage. Our products work well internally, but people’s imaginations are limited. If we open NoCode to a wider range of scenarios and allow more people to use it, it may bring more possibilities, which will create value not only for us but also for the entire industry.

Meituan has always focused on retail plus technology. In addition to relatively hard-core technologies such as automatic delivery vehicles and drones, we are also iterating quickly on soft-core technologies such as AI models.

Compared to some AI code generation tools on the market, I think we open it up to allow everyone to participate in AI. We will continue to invest in the direction of retail plus technology, and have made some good progress.

3. What are the capabilities of NoCode? Do you have any code benchmark results you can share?

Cheng Datong: We don’t list our code on the external list. In internal testing, we can see that Code Agent has a relatively high score on SWE-Bench (Note: a large model capability evaluation dataset based on GitHub). The industry SOTA level on the Verified project is around 65% after fine-tuning. Our score can reach more than 50% without fine-tuning, which is almost at the top 5 level.

We are more concerned about the actual generation effect. In internal practice, we can see that the usage rate of NoCode is very high, and the number of non-technical users is three times that of technical users.

We have a lot of practical experience in AI code generation. Compared with startups, we already have a considerable number of "customers" - Meituan's first-line BD and sales, and even some B-side merchants, who are already very proficient in using it. In addition, our experience and accumulation in intelligent agents also promoted the construction of the technology stack.

We hope that the application of AI technology can be more popular, with more automation from demand, debugging to environment deployment, so that application construction can be achieved from zero to one.

The technology behind NoCode and Dev Mode

4. Please briefly introduce the general framework of NoCode. What is the basic model supporting NoCode?

Yu Chao: NoCode is a large-model intelligent product built around infrastructure, runtime sandbox and large models, with multiple AI models collaborating.

High-resolution image link: https://nocode.host/6281d/lp1779rwn2cpl5lms 

Generally speaking, the overall architecture of NoCode is divided into three layers, namely the infrastructure and model layer, the runtime sandbox and baas layer, and the Agent application layer. At the infrastructure level, it includes K8S large-scale cluster scheduling computing power, distributed file systems to provide code storage capabilities, Ingress gateways to implement website routing, deployment systems to provide deployment capabilities, and image retrieval services to adapt to web page images. Further up is the runtime sandbox layer, which builds images based on specific technical frameworks (such as react, cloud IDE), accelerates cold starts through pooling technology, and implements real-time page rendering through hot updates. Finally, there is the Agent application architecture, a large-scale main inference model, plus some small models for vertical scenarios, to build a code agent including plan, context management, and tool sets in code scenarios.

After switching to Dev Mode, the user view changes to a complete Agentic IDE - CatPaw, which is suitable for professional users to perform more refined control. CatPaw was launched in November 2022. At that time, it was called MCopilot. Later, we thought that the word Copilot was not a good word, so we changed the name to CatPaw. Currently, the penetration rate of Meituan's internal research and development exceeds 90%. This part of the technical architecture has more complex details, and there will be opportunities to share it separately later.

High-resolution image link: https://nocode.host/6281d/snpwps0uq1nje77bg

5. Among these, what innovations has Meituan’s engineering team achieved, and what are its core technologies (such as model training and tuning, data sets, etc.)?

Yu Chao: We have specially built a 7B parameter apply-specific model to solve the problem of slow generation of large-size models. This dedicated model is based on Meituan’s self-developed code base and is trained with Meituan’s internal real code data and some synthetic data. Its task is to generate full code based on source files + modification plans. After evaluating its generation indicators offline, we conduct A/B experiments online to measure the final end-to-end operation effect of the entire product, and continuously optimize and iterate. At present, more than a dozen versions have been iterated.

Apply dedicated models. We have a dedicated optimization team that continuously optimizes algorithms and engineering, and can achieve inference of 2,000 tokens per second without losing accuracy. Moreover, we do not use the most advanced GPU card in the industry. If a large-size model is used to output code, the generation speed may only reach dozens of tokens per second.

In fact, in addition to the Apply model, we have also trained many small models on our own, such as the earliest code completion model, Embedding model, Rerank model, etc. Their effect in the architecture is to improve the reasoning speed without reducing the indicators, so that it can be fast and good. Relatively speaking, using a large-size model seems to waste tokens.

For model iteration, the quality of data and the speed of experimental iteration are crucial. We built training sets and evaluation sets based on Meituan’s internal data and synthetic data, and added manual proofreading and review. In order to improve the effect of vertical scenarios, we conducted a lot of offline evaluations and online A/B experiments. We not only conduct vertical evaluations on small models, but also evaluate the entire end-to-end link of the product. Our goal is not to hit the Benchmark rankings, but to continue to fine-tune according to the needs of developers and non-developers in the Meituan ecosystem.

6. Opening up to the outside world means the possibility of large-scale application. In terms of product experience, what optimizations has Meituan made to NoCode?

Yu Chao: The ultimate technical optimization brings the ultimate product experience. For example, real-time code rendering actually requires the computing power of the server. The best situation is that it is like local development. After changing the code, I can see the output effect immediately, and then I can quickly iterate.

During the conversation, each NoCode work has a runtime sandbox behind it, which has the environment, dependencies, IDE, etc. installed in advance. It can quickly help you update the code to the container in real time, and quickly see the effect after hot update.

Of course, this will also bring another problem, the overhead of container resources is very high. We have specially designed a mechanism for recycling containers, combined with stateless design and container pooling, so that you can start the container in seconds the next time you open it.

In order to combat the instability of large model output, we also designed Visual Edit. You can select a part of the web page to make local modifications, such as changing a text or a background image, without relying on a large model. It is fast and accurate.

After each conversation, we will generate a version with page screenshots by default. By browsing the history with pictures, you can easily review each change, making it convenient to switch versions or roll back with one click.

If you have used NoCode, you should have noticed that the images in the web pages generated by NoCode are very suitable for the scenario of generating web page context. This actually involves image search, and we also have a team to optimize this service. Specifically, we will combine the context of the web page, use keywords and semantics to search for images, and then place them on the web page. Appropriate images are very important for the overall beauty of the web page.

Finally, in order to adapt to the demands of more professional users and solve the collaboration between non-professional users and professional users, we designed Dev Mode. Dev Mode is almost like moving the more professional version of AI Native IDE within Meituan to the browser to achieve more refined control. After some iterations of using NoCode, if non-professional users find that they can't make changes, they can directly share the link to professional users, who will switch to Dev Mode for rapid iteration. Non-professional users can refresh to see real-time effects and achieve collaborative creation.

These are just a few examples. We still have a lot of optimizations on the way, and we welcome everyone to give us more suggestions in the community (feedback.nocode.cn).

7. How to make AI follow the instructions given by humans more accurately and write useful code? What challenges did Meituan’s engineers overcome in the process of building NoCode?

Yu Chao: The ability to follow instructions mainly tests the model's reasoning ability. In addition, it is also necessary to provide enough unambiguous context to the model. In the NoCode scenario, it is also necessary to select the front-end technology framework that has the best model output effect. For each model, we spend a lot of time adjusting the system prompt, such as whether the model is proactive enough, the code specifications used, whether the tool description is accurate, whether the format is optimal, etc. We will also support some user-defined system prompts in the future, and you can formulate some to suit your personal habits. At the front-end technology framework level, we chose a combination of react+tailwind+shadcn+vite. In addition, the code output by the model sometimes fails to compile, and we will automatically give the compiler error to the model for repair until it can be successfully rendered. The repair strategy here will also have special optimization and iteration by the team.

8. When people build projects in real production environments, they usually need large model tools to maintain consistency and have long-term memory capabilities. Does NoCode have such a mechanism and how does it do it?

Yu Chao: If there are not many code files, you can directly put all the codes into a window. After exceeding the context window of the model, we will use RAG to achieve knowledge enhancement on the one hand, and provide the model with some tools for code retrieval on the other hand. We have specially trained a series of search models for RAG scenarios. For example, how the embedding model vectorizes the data set, the rerank model performs rearrangement, and so on. Every task and every detail in this has been specially trained and evaluated, and the extreme is achieved on a single point, and then it is connected in series for combined experiments. As the reasoning ability of the model is enhanced, after more tools are directly trained into the model, the model can dynamically plan subsequent actions based on the current context, such as retrieving the code repository by keywords, or listing the directory tree structure of a directory, thereby enhancing subsequent reasoning.

AI enables everyone to write good applications

9. We know that using AI code tools often requires multiple generation and adjustments before practical code can be output. For non-professionals, how do we know whether the code generated by AI is useful? How does NoCode help people iteratively optimize the code?

Yu Chao: You need to iterate continuously, input your ideas first, and see what results it will produce. Many of our internal employees were slightly inexperienced when they first used NoCode, and the results they generated were completely different from their expectations. But they gradually figured out the AI's temper during the conversation. In my opinion, each model has its own temper, which is "modelability". In the process of using it, you will find out what kind of instructions are the most stable and easiest to follow for it. Gradually form your own method experience.

The learning speed should be very fast, because you don’t need to learn a lot of boring codes and understand the grammatical structure and rules. You only need to improve your description and adjust the output of the model.

When using NoCode, if you find that the model output does not meet your expectations, you can easily roll back by clicking the mouse in the visualized version history on the right side of the page. After rolling back, you can add more information or change the description and let NoCode help you generate a new version.

We will also release some practical cases to fully present the dialogue process of each work, and there will also be a community for everyone to share experiences. I believe that the learning threshold of such a product is much lower than that of traditional code, and you may be able to find an efficient implementation method in one day.

10. How will the workflow change when using NoCode to write applications?

Cheng Datong: It mainly depends on the role of the user.

For a product manager , if I put forward a requirement in the past, I might need to find a R&D person to make a demo, then go through the design, and then review it, which might take a week or two. Now I can use AI tools to directly make a prototype. If the developer thinks it is OK, he can modify it directly on the prototype and carry out deeper development. If you have strong hands-on skills, you can even use AI tools to make further modifications yourself.

Many of our front-line business personnel can quickly build some urgent small projects in this way to achieve a closed loop of business.

The third category is the completion of full-stack capabilities . If I write a backend and need to write a frontend myself, I can use NoCode to help me. As a development engineer, it expands my technology stack.

There are also data analysts who used to use Excel extensively and needed to know VB programming. Now they can directly analyze through natural language, and AI can help them fill in data and generate charts.

It can be said that everyone is changing, including the workflow and collaboration mode in the entire chain. We have done some research and found that whether you know programming or not, the way of working can change.

11. What efficiency improvements have people achieved by using NoCode within Meituan? Are there any more specific examples?

Cheng Datong: At Meituan, many employees who are connected to overseas businesses use NoCode to build software for learning foreign languages. They design products according to personal learning habits and needs, introduce third-party tools, or engage in collaborative learning. Some people even use NoCode to write lottery programs.

Our HR team and administrative team have also changed. For example, the red envelopes that the HR team needed to send out during the Chinese New Year last year were made by some students using NoCode. With the assistance of R&D, they helped 100,000 people in the company receive red envelopes. They only took one to two working days to complete this task.

In the past, data analysis in the data team required the use of professional software, many of which were based on the underlying big data warehouse. However, such tools are general-purpose, and for customized requirements, you need to submit requirements to the dedicated software team. Now people do not need to submit requirements, they just need to use NoCode to make an on-end data analysis tool, which basically only requires 30 to 40 rounds of conversations and a morning.

With structured data and NoCode, I can quickly meet my needs.

12. What percentage of the code submitted by Meituan every day is built by AI code generation tools. How much can the development time be shortened by using natural language to create production-ready applications?

Cheng Datong: In the first quarter financial report this year, it was mentioned that Meituan used AI to generate code accounting for 27%. This number is still increasing. Currently, the incremental code generated by our internal AI tools accounts for 50% of the incremental code in the warehouse every week. The statistics for the first quarter have not yet included NoCode.

We believe that (using AI development tools) can bring 30-50% efficiency improvements internally.

Project background and future prospects

13. Where did the NoCode project come from? Can you briefly introduce the history and development team behind it?

Cheng Datong: Our team has been working on AI Coding code agents for two and a half years, but the launch of the NoCode project has only been six months. It started in October 2024. It was initially an internal hackathon project with only three students, and now there are less than ten core technical personnel.

In November last year, more than a month after the project started, we suddenly discovered that this project could be popular, so we quickly ramped up the project.

Of course, there is definitely mutual dependence and cognitive iteration. I think the fact that we were able to quickly produce this product has a lot to do with our previous accumulation.

In the continuous development of AI products, we found that we must have rapid iterations, and you can’t be late by even a minute. In addition, in the attempt of this new direction, new AI products must pay attention to details, especially in the vertical category.

We did things that the model base could not do, and many of the details were polished by us, such as the link to the cloud disk during deployment, and the problem of calling tool dependency versions in model knowledge. All of these are optimizations on the link.

Yu Chao: NoCode releases a new version at least once a week internally, and sometimes twice a week. After converting the internal version into an external version, we will continue to iterate at this pace.

14. On the other hand, for people who use AI, using big models is an amplification of existing professional knowledge. What new requirements will the new forms of collaboration between humans and AI in the future put on people?

Cheng Datong: There are many things that need to change. I think everyone needs to continue learning. For participants, you have to read papers, understand the algorithms and engineering problems from the backend to the frontend, and you have to use many products.

For users of AI products, it is important to get started. You need to enter your first query and be patient. It is not the era of AGI yet, but more and more things can be done by agents automatically. You need to learn to deliver results quickly in this human-in-the-loop environment, not just improve efficiency.

If you have more professional programming knowledge, you can also try more professional development tools.

I think AI programming will become like the relationship between "car and driver" in the future - thirty years ago, everyone thought that driver was a profession, but now basically everyone can drive. With the emergence of autonomous driving in the future, there will even be no need to drive. Programming skills will become very universal.

15. What’s next? Can you tell us about the future development direction of NoCode? Will Meituan launch more productivity applications?

Cheng Datong: For NoCode, we will continue to improve stability and experience in the near future and continue to optimize the model. In the longer term, the development direction may be to open up the automation of AI development from non-professional to professional. We hope to provide a better development environment and explore the IDE field.

We may release the development tool "Dev Mode" in June to implement a more professional IDE and build full code and compilation capabilities.

Meituan demonstrated the Dev Mode that is under development, providing more complex capabilities for more professional needs.

We don’t know what kind of impact NoCode will have on the market, but I think we should create as much impact as possible. Our goal is to make good products with a pragmatic attitude, focus on results, and focus on product capabilities.

© THE END