Two and a half years after Light Year (Guangnian Zhiwai) was folded into Meituan, the team now known as GN06 is still working toward the same goal: identifying artificial intelligence-driven growth opportunities beyond the company’s core businesses.
In March, that effort took a more concrete form with the public beta launch of Tabbit, an AI-native browser.
Why a browser? As large companies race to claim ground in AI, the industry’s buzzwords are “lobster,” a colloquial term for OpenClaw, agents, and tokens. A browser is neither the most fashionable format nor an obvious extension of Meituan’s existing businesses.
Liu Jiong, who leads GN06, sees it differently. There may be many ways to build AI applications, he said, but he wants a browser that can do work on behalf of nontechnical beginners and give people back time.
“Plenty of students, knowledge workers, and creators spend most of their day on computers, and most of that computer time is spent inside a browser,” Liu said. “If AI can genuinely help people there, the room for imagination is enormous.”
Tabbit’s latest public beta version is 0.25, still short of version 1.0. Yet the team said retention has stayed above what it considers a strong industry benchmark, and it continues to receive notable user feedback. According to the team, one cross-border e-commerce user built eight automation “skills” covering an entire workflow. A Tsinghua University student used the product to organize large batches of academic papers. A blind lawyer has used its accessibility features and agent capabilities in daily work, then recommended it within his community.
Liu describes the team’s goal for Tabbit in simple terms: understand users’ pain points and build close to real-world scenarios. If the browser can save users even half an hour a day on repetitive work, helping them leave earlier or freeing up energy to pursue their interests, then the product has value.
The following transcript has been edited and consolidated for brevity and clarity.
36Kr: We’ve heard you wanted to build a browser quite early. Why?
Liu Jiong (LJ): We saw that many people open a chatbot input box and do not know what to type, or how to begin, almost like introverts in MBTI terms. From my perspective, for the vast majority of people, the time spent looking will always far exceed the time spent writing. Whether it is short videos, articles, or news, the time you spend browsing and reading is greater than the time you spend producing output.
So what is something users will keep looking at? Short videos are certainly one, but that is not something we can do.
The other is web browsing. Huge numbers of students, knowledge workers, and self-media creators spend most of their time on computers, and most of that computer time is spent inside browsers. That is where they search for information, read content, and complete writing tasks. There is a natural, real demand there.
More importantly, bringing AI into a user’s existing work environment is much easier than asking that user to adopt something entirely new.
36Kr: Do you think a browser is closer to daily use cases than a chatbot?
LJ: I would say it understands user needs better, because a browser has your most complete context.
Let me give you a real example. One of our colleagues was reading an internal company document that used the term “CPL.” If you copied that term into any standalone AI chat box and asked, “What is CPL?” it would invent eight generic business explanations. But in Tabbit, if you highlight it and ask the question, the AI will look at the full context of the page you are viewing and tell you, “Based on the context, this should be CPI. It was written incorrectly.”
That is where the browser has an overwhelming advantage. It knows what you are looking at, and it understands your context.
36Kr: During Tabbit’s planning process, what was the biggest internal debate?
LJ: What happens if Google Chrome puts AI into the browser?
36Kr: Isn’t Chrome already doing that now?
LJ: Yes, so that question has always been there. I am more optimistic about it. In 2023, people thought many things had no chance of being transformed by AI.
Take programming. At the time, Microsoft’s Visual Studio Code was the default tool for almost every programmer. GitHub belonged to Microsoft, Microsoft had launched Copilot, and it was also an investor in OpenAI. Everyone thought no one could beat Microsoft in programming. But today, you have a whole wave of new-generation products, including Cursor, Codex, and Claude Code.
The same applies to office software and browsers. The companies behind leading products will definitely move to make those products AI-enabled. But they also tend to carry more legacy burdens, and their AI strategies are often the opposite of ours. Their first objective is to put their own large model into the application. Google’s thinking, for example, will definitely be to put Gemini into the browser and nothing else. OpenAI will do the same with ChatGPT in Atlas.
36Kr: So your argument is that your approach is more open?
LJ: Our motivation is different. Their thinking is, “I have a model, and I want to build a very good entry point to bind users to it and sell that model.” We have connected leading models from multiple companies, and users can switch freely.
36Kr: Have you faced any resistance on model choice?
LJ: Meituan has its own model, LongCat, but not for a single day has the company said that only LongCat can be used. Other model providers have also been very open. They are very willing to get more prominent placement and resource support inside our product.
Behind that is the fact that Tabbit has many user scenarios of its own. Some vendors will say, can we work together to build certain datasets, improve success rates in some agent scenarios, or jointly fine-tune models for specific use cases?
36Kr: What is the downside of other players binding deeply to a single model?
LJ: This is a highly competitive space right now. Which model is the strongest today? The answer keeps changing, and different scenarios are suited to different models. When a company can use only its own model, it takes on the risk that its model may not be the best choice.
From 2023 until now, no company’s model has remained state of the art the entire time. Leadership keeps shifting. I do not think that competitive picture will change much over the next two or three years. In the early stage, openness matters a lot.
36Kr: How much does offering multiple model choices affect user experience?
LJ: If 90% of users all had a strong preference for a single model, then maybe our product design itself would be wrong. But the data shows users are very clearly switching models. Looking at our model call volume, even the largest single provider accounts for less than 30%. That is already quite dispersed.
We have also received feedback from users who said they previously had to pay OpenAI, Google Gemini, and Claude USD 20 each, or USD 60 in total, just to use the best model for different scenarios. With us, even if we charge in the future, maybe they can access all of them for USD 20. For that user, it is also a cost saving.
36Kr: How has the strategy of building the “most open” AI-powered browser worked so far?
LJ: We officially launched public beta on March 2. It drew in an initial group of users, and we received a lot of positive feedback, along with ongoing word-of-mouth growth. Judging from the data, retention overall is very strong for a productivity tool.
36Kr: For the broadest group of users, how high is Tabbit’s barrier to entry?
LJ: There is no barrier. The switching cost is zero. You just click import, and your Chrome bookmarks, open pages, and logged-in accounts all move into Tabbit seamlessly. The experience is also fully compatible with Chrome. Whether you used Chrome, Edge, or another Chromium-based browser before, you can get started in a second. All the models are ready to use out of the box, without the complicated setup and API key management that lobster-style tools often require.
My own personality is relatively restrained. I really dislike all those floating plugin bubbles on webpages. I am almost obsessive about it. That has pros and cons. The good side is that our pages are very clean. The bad side is that users sometimes do not find some of the good features. So when it comes to guiding users through useful functions, what each feature can do, and how it works, we clearly have not done enough. Next, we will work more on the operations side and use strong case studies to show users what is possible.
36Kr: Such as? What usage scenarios have impressed you most so far?
LJ: Cross-border e-commerce users reacted very quickly. They quickly realized agent capabilities could help them work more efficiently. One user created eight e-commerce “skills” with Tabbit, covering an entire business workflow, from data scraping and price comparison analysis to interpretation of operations dashboards, all completed automatically.
Students are another very interesting group. Some Tsinghua students, when writing papers and needing to read large volumes of literature at once, use Tabbit’s AI capabilities to summarize all of it simultaneously, interact with local files, and sort and organize the material, significantly improving efficiency.
Our company’s HR team also created a “skill” that automatically reviews resumes on recruiting websites according to specific conditions. Once it was written, it could be shared directly with other HR colleagues, who could use it in one click. That also saved a great deal of time.
There was also one user who surprised us in particular: a blind lawyer. In the past, it was very difficult for him to fill out forms on webpages or shop online. But through Tabbit’s agent capabilities, he can issue voice instructions, and the AI can read screens for him, compare prices, and even help him understand the plot of foreign films. He has recommended it to other blind friends, who are also trying to use it to prepare lessons for students at schools for the blind or to work on startup projects.
36Kr: If Tabbit achieves wider adoption, what do you think will drive it?
LJ: We hope it happens through all kinds of useful and interesting “skills.” Different skills can unlock different possibilities for different groups. Even within the same group, there may be a special bundle of skills. Once you open that bundle, you realize that your daily work has all become simpler.
36Kr: And there will also be uses you did not anticipate.
LJ: Exactly. What we are building is highly practical, but there are also many playful users who like to experiment. In fact, browsers are inherently very open. Apart from the core framework, 90% of the space is made up of webpages, so you can do almost anything.
After we opened up scripting capabilities, for example, one user thought a webpage looked too plain and used AI to generate a script that turned all the chat bubbles pink. Another made a little dog run across the page while typing, then sit down when typing stopped. Others created one-click scripts to add dark mode to webpages or automatically generate tables of contents for long articles.
We are now providing a foundation that lets users get creative and hand that creativity back to them.
36Kr: If bigger companies pour huge resources into making chatbot-style entry points the default gateway in the AI era, does building a browser still matter?
LJ: Yes. For many users, it does not necessarily mean they will rely on a single product in every scenario. The browser market is large enough. Chrome has roughly four billion users. If we capture just 1% of that, that is 40 million users. Even one-thousandth of that would still be four million users.
I think what we are doing and what those chatbots are doing are two different things by definition. They may be general-purpose entry points that can do many things. Meituan is also doing that. Xiaomei, for example, is a general-purpose lifestyle gateway. Our team is building more of a workstation for knowledge workers, students, and office employees, a tool.
36Kr: Is the overseas market a useful reference point for you now? For example, how does competition between Chrome and newer browser challengers look today?
LJ: There are several leading players overseas, and we can look at them one by one.
One is Atlas from OpenAI. But OpenAI’s front is stretched too wide. It was doing Sora 2 and a whole range of other things, and I do not think fighting on every front leads to a complete win. So narrowing focus is a good thing. We are watching Atlas closely. It has been upgrading and iterating continuously, and I think it is definitely in the core top tier.
Dia has a user base because of its strong UI design and the foundation built by Arc. Comet also has relatively strong agent capabilities. Competition is still in its early stages.
Then there is Chrome. Since the end of last year, from what we have observed, its pace of iteration has been accelerating. That includes its integration of Gemini, as well as the vertical tab bar, something it had not introduced in ages and is now developing quickly.
36Kr: What is the fundamental difference between a vertical sidebar and a top tab bar?
LJ: If you have many tabs open, a top tab bar turns into a row of small dots and icons, making it difficult to identify each page. The main advantage of a vertical sidebar is that it shows both the icon and the title. At a glance, you can see which page is which. This has a meaningful impact on user experience, and many users prefer it.
36Kr: From a technical standpoint, is moving the bar from the top to the side really that hard?
LJ: Very. In fact, very few browsers have built a native vertical sidebar in C++.
Broadly speaking, there are three main approaches to developing a browser. The simplest is to build the product as a web app, effectively a webpage within a webpage. Development is fast, a small team could complete it in a matter of months. However, embedding a webpage inside another webpage can cause performance issues. The system may freeze when opened, and some AI-powered browsers can drain a laptop battery within an hour. Lag and overall performance are the main constraints.
The second route is represented by Dia, Arc, and Atlas. Their approach is based on CEF, or Chromium Embedded Framework, which is also an embedded layer extracted from Chromium. But unlike Electron, CEF can use native system interfaces, so performance is better and the experience is smoother. On macOS, they use the SwiftUI framework to deliver a strong interactive experience. But that comes at a cost: because that native UI is deeply tied to the operating system, it cannot be directly ported to Windows.
In the end, we chose the hardest route: building an entire stack in C++ on top of Chromium’s overall architecture. It is slower at the start, but the payoff is better. When we were researching early on, we looked at these routes and the future problems each one could create. Many features may not look difficult from the outside, but the technical strategy and interaction design behind them require a great deal of thought.
36Kr: Is Tabbit now well adapted to Windows?
LJ: From day one, we decided to support Windows, because in every country there are far more Windows users than Mac users. User feedback has reinforced that decision. Since Tabbit entered public beta, the number of Windows users has already exceeded the number of Mac users by a wide margin.
36Kr: At this point, do you believe the product’s overall direction is the right one?
LJ: My view is that, for an AI application team, the essential task is to understand user context and use cases, then build close to those scenarios. Whether the final product is a browser or not, I am open.
We call it a browser because it is easy for users to understand and represents a clear use case. White-collar workers spend five to eight hours a day in browsers. Of course, I am aware there are critics.
36Kr: What kind of feedback or criticism have you heard?
LJ: Some people think the browser is the wrong path. Some peers have tried it before and then abandoned it. There’s plenty of such feedback.
If we compare the ambition of using AI to reshape the world to China’s college entrance exam, then the industry remains at an early stage. The field has developed only over the past two to three years. Many of these judgments are premature, and I do not place much weight on them.
36Kr: Are there any more fundamental criticisms?
LJ: At a deeper level, the narrative is simply not as compelling as OpenClaw’s. But why should I care about that? I am not fundraising. I do not like arguing with people. What I care about is reading user feedback and speaking with people who have direct experience with the product.
36Kr: People have suggested that agents will browse the web on users’ behalf in the future. What do you think?
LJ: I believe agents may take over many tasks. But as a person, I am curious and want to explore. I will still open Dribbble to look at designers’ work, watch videos on Bilibili, and read articles on WeChat official accounts. I take in information, think independently, and write. People have needs both to consume and to create content. I do not think browsers will disappear, but AI can improve how we browse and execute tasks.
Likewise, our goal with Tabbit is simple: understand users’ pain points and build close to real-world scenarios. If it can save users even half an hour a day on repetitive work, helping them leave earlier or giving them more energy to pursue their interests, then the browser has demonstrated its value.
KrASIA features translated and adapted content that was originally published by 36Kr. This article was written by Ren Cairu for 36Kr.