Permanent free access, Meituan AI Browser is here! Over 10 top-tier models help you get work done, with an extra Agent plugin included.

On June 9th, it was reported that the Meituan GN06 (formerly Yuanlight Zhiwai) team officially released the AI browser Tabbit V1.0 and promised that the core functions would be permanently free.

Since its public beta launch on March 2nd, Tabbit has been available for exactly 100 days. It has been iterated weekly, with a total of 12 versions, and has received a lot of positive feedback from users, such as "the most beautiful browser on Windows", "a very practical tool product", "a low - threshold and safe way to use top - tier models", and so on.

Liu Jiong, the person - in - charge of GN06, revealed that Tabbit has been in good condition over the past 100 days.

This AI browser inherits the usage habits of traditional browsers and integrates the capabilities of "large - model Q&A + search + Agent".

It can execute complex Agent tasks according to user instructions, operate web pages, generate PDF, PPT, Excel, and Word documents. It is compatible with mainstream Agent Skills and MCP tools, and can freely call more than 300 ready - made Agent skills in the Magic Trick Square, acting as an all - around assistant to help users with their work.

It's worth mentioning that Tabbit not only supports the invocation of more than 10 top domestic large models (including DeepSeek, LongCat, GLM, Kimi, MiniMax, Doubao, Qwen, etc.), but also has the multi - model selection function, which allows 5 models to answer the same question simultaneously on the same page and enables AI to make comparative summaries.

In Tabbit, it is very convenient to activate the AI. You can open a new tab dialog box to ask questions directly, or open the sidebar dialog box from the "Chat" button in the upper - right corner, or select words or take screenshots on the web page to ask precise questions.

Moreover, Tabbit will remember the preferences, backgrounds, and important information repeatedly mentioned by users in historical conversations and actively adapt to users' habits.

For example, Tabbit can do the following things:

• Refer to 10 pieces of materials and generate a college major recommendation plan for the college entrance examination that matches colleges and majors through several rounds of conversations.
• With a link to an online course video and a sentence, generate a set of materials with core concepts, theorems, methods, and self - test questions in 3 minutes.
• Create a simple web game casually.
• Verify the information source of a link and analyze the rhetorical tricks.
• Select words on a PDF to view definitions, examples, and word roots, and ask questions as you read.
• Generate an outline and a PPT based on the uploaded materials and screen - reading information.

Liu Jiong said that Tabbit's slogan is: Leave your work to Tabbit and keep your time for yourself.

This week, Tabbit will gradually push the update of V1.0 and promises that the standard version will be permanently free.

The weekly free quota allows users to use the latest AI model for about 1000 conversations, generate 50 illustrations, let the Agent automatically complete 10 tasks, or write 100 summaries of 10,000 words or articles of 1000 words.

If the free quota is used up, users can continue to use an unlimited and cost - effective model.

In the future, Tabbit will also offer a professional version with 10 times the usage, priced at 9.9 yuan per week, along with more core functions.

"For the price of a cup of Luckin Americano, you can have more than 10 top domestic models work for you," Liu Jiong said.

The trial version of Tabbit is now available on mobile phones. The iOS and Android versions have been launched, and the HarmonyOS version is under development. The functions are being gradually improved, and currently, bookmark synchronization and ordering takeaways are not supported.

For users who participated in the previous public beta, Tabbit will give away an exclusive commemorative card for "Pioneer Exploration Users" and offer professional - version privileges based on the registration time and usage days.

01.

Behind the Reconstruction of the Browser:

The Perfect Entrance in the AI Era

Why Develop an AI Browser?

Liu Jiong explained that the framework of traditional browsers has been fixed, but the complexity of web pages themselves is increasing rapidly. Today, white - collar workers spend more than 66 hours per week on browsers on average, and tasks such as office work, collaboration, document processing, meetings, emails, and data analysis are all completed in browsers.

In his opinion, browsers are not outdated; instead, they have become even more important because of AI. The applications, documents, and tools generated by AI are all based on web pages.

There are already desktop intelligent agent tools like Lobster and Codex in the market. Why do we still need an AI browser?

Currently, the way we use AI tools is often fragmented. We need to open multiple software, search for information in one window and then switch to another window to ask the AI. Operations such as copying, pasting, and uploading files will consume extra attention and time, and there will also be information loss.

An AI browser naturally has access to the tabs you open, the articles you bookmark, and the web pages you are browsing. All these core contexts do not need to be transferred. The AI is always there, sharing the same environment with you, knowing what you are looking at, and can act directly based on the context in your browser.

This enables the AI to better understand and more efficiently participate in your work.

However, many users have already developed habits with existing browsers, and their bookmarks and plugins are all in the existing browsers. Why should they switch to a new browser?

In response, Tabbit has upgraded functions such as full - intention input, personal knowledge base, memory source, AI perception and operation, and ongoing tasks on the basis of the traditional address bar, bookmark folder, history record, tabs, and tab groups. It upgrades the underlying intelligence based on users' existing browser habits.

Its bookmark function not only saves a URL link but also saves the entire content of the web page. Even if the web page becomes invalid, you can still view it, and the AI can directly quote the full - text content of the article in subsequent conversations.

02.

Comprehensive Evolution of Model, Personalization, and Experience:

Upgrade of General Agent Capabilities and Expansion of the Magic Trick Ecosystem

Compared with the previous public beta version, the model, personalization, and experience of Tabbit V1.0 have all been comprehensively evolved.

1. Model: Connect to more than 10 top - tier models and select multiple models to execute tasks simultaneously

Tabbit supports free switching among more than a dozen leading domestic AI models (MiniMax M3, DeepSeek - V4, Kimi K2.6, GLM - 5.1, Doubao - Seed - 2.0, Qwen3.5, LongCat - Flash, etc.) for users to choose according to their needs.

The "Add Reference" option contains all the opened tabs. By entering @, you can reference the currently opened tabs, tab groups, bookmark folders, or upload local files, and use the web - page content as context for AI tasks.

Liu Jiong demonstrated searching for "WWDC" on Xiaohongshu and using the "Export Xiaohongshu Comments" magic trick to capture and organize the comments in a post.

Then, by entering "Organize the content of the CSV file into a Markdown - format material and analyze the core ideas", Tabbit will call the AI to complete the task.

In the "More Modes" option, the "Multi - model Selection" function allows you to select 5 models simultaneously to compete and compare their answers to the same question.

It can also analyze and summarize the answers of these models.

2. Personalization: Controllable Memory and Reusable Magic Tricks to Improve Agent Task Completion Rate

The Agent task mode has stronger capabilities, more scenarios, and enhanced security.

(1) Upgrade of the running environment (Harness): Page snapshots, screenshots, DOM understanding and execution; self - developed browser operation tool set; multi - step task planning and process intervention; comprehensive enhancement of web - script generation.

(2) Editable and Deliverable: The success rate of tasks such as online document and multi - dimensional table processing has increased by 22%; generate PDF/PPT/Excel/Word, etc.; conduct multi - source information research; be compatible with the mainstream Agent Skill/MCP ecosystem.

(3) Cloud Execution without Polluting the Local Computer: Agent orchestration and code execution are carried out in the cloud. Local files and folders are accessed according to authorization, and there will be no need to install a bunch of messy things or create a large number of files on your computer.

In Meituan's own benchmark, from Tabbit V0.23 to V1.0, the success rate of web - operation tasks has increased from 61% to 70%, and the overall score has increased from 65.0 to 94.3.

According to Liu Jiong, Tabbit has the complete context of users. On the basis of the web pages, bookmarks, Chat conversations, and task conversations in traditional browsers, it has filled in the three pieces of context puzzles of cloud MCP, local directory, and memory.

The Beta technical preview version has long - term memory, which