WeChat "Xiaowei": A Complete Tour of Its 12 Entry Points, Baking AI into Every Information-Consumption Scenario
WeChat's native AI assistant "Xiaowei" has entered gray-scale testing. Built on WeChat's in-house WeLM model, it offers smart retrieval and processing across 12 entry points spanning chat history, Official Accounts, Channels, documents, and more.


WeChat "Xiaowei": A Complete Tour of Its 12 Entry Points, Baking AI into Every Information-Consumption Scenario
WeChat's native AI assistant "Xiaowei" has entered gray-scale testing. Built on WeChat's in-house WeLM model, it offers smart retrieval and processing across 12 entry points spanning chat history, Official Accounts, Channels, documents, and more.
WeChat's native AI assistant "Xiaowei" is rolling out in gray-scale testing. Its biggest difference from mainstream chatbots is this: chatbots extend application scenarios inside a chat box, while WeChat's approach is to weave AI naturally into its existing high-frequency scenarios — chats, Official Accounts, Channels, and documents. Xiaowei's main model is WeLM, developed in-house by the WeChat team, with some answers drawing on DeepSeek. After a full day of intensive hands-on testing, we found 12 entry points inside WeChat, which fall into four categories.
Category 1: The green eyes in the top-left corner (main entry)
Once you get beta access, a pair of green eyes appears in the top-left corner of the WeChat home screen; tapping it opens Xiaowei's main interface. Voice input is the default at the bottom and can be switched to text, and the "+" button in the bottom-right corner lets you upload photos, use the camera, or attach files.
This entry serves as the "master gateway" to Xiaowei and covers at least the following:
- Social information: Moments updates, Official Account updates, article search and summarization
- Native WeChat actions: sending red packets, checking to-dos, system settings
- Local life services: navigation, ride-hailing, and other Mini Program capabilities for nearby services
- AI creation: image generation and editing
- AI app generation: creating mini tools from a single sentence
Ask "catch me up on what's new in my Moments over the past couple of days" and it organizes posts by topic (AI and agents, autonomous driving, quantum computing, and so on); ask "where has everyone been traveling" and it sorts travel updates into groups like outdoors and nature, city business trips, and overseas travel. Ask "what articles has the Geek Things Official Account published lately" and it returns article cards with summaries; follow up with "which one is worth reading" or "summarize the key points for me" to dig deeper.
For anything involving payments, contacts, or sending messages, Xiaowei always requires a second confirmation — an inevitable caution from a platform serving over a billion users and an entire payment system.
Category 2: Chat interface entries (with context)
Three entry points hide inside the chat interface: long-pressing a chat message, the "+" button in the bottom-right of a one-on-one chat, and the "+" button in the bottom-right of a group chat. Their value is that they carry the chat context — the home-screen entry cannot directly access a specific friend's or group's content, while invoking Xiaowei inside a chat lets it understand the current conversation.
Ask directly "what was that remote control ID he sent me earlier" and Xiaowei will dig the record out of the chat history, give you the time and verification code, and ask which instance you meant. In group chats you can:
- Look up who mentioned me
- Filter for important messages
- Draft replies based on the group conversation
- Summarize the group's topics
- Organize to-dos buried in the chat history (shown as to-do cards)
Compared with traditional keyword-based search, this context-based questioning is far more direct, and especially practical for users drowning in work-group messages.
Category 3: Official Accounts and Channels (summarize before you watch)
Both Official Account articles and Channels support invoking Xiaowei at share time, which amounts to adding a "preview summary" before you read or watch.
- Official Accounts: it distills an article's main thesis, key events, and core information. Ask it to summarize an article about WeCom's AI agent, for example, and it will outline the product form and core scenarios.
- Channels: send a conversation video to Xiaowei and it can summarize the key takeaways (such as someone's three reasons for joining a company, or the problems of pairing models with products in AI's second half).
This changes how you read — instead of skimming headlines and clicking in, you now let Xiaowei screen things first and then decide whether to open the original.
Category 4: Document entries (images and the office trio)
Document-type entries cover images, PDF, TXT, Word, PPT, and Excel. Long-press an enlarged image to ask Xiaowei about it; for the other file types, tap share after opening the file to invoke it.
These entries skew toward workflows. Huge numbers of files move through WeChat every day; previously you had to download and open them or forward them to a computer, but now, when you just want a quick sense of what a file says, you can ask Xiaowei right inside WeChat:
- Image understanding: long-press an image to ask about its contents directly, turning the image's key information into text
- Document processing: PDF, TXT, Word, PPT, and Excel files can all be handed to Xiaowei to read, summarize, and organize
What is certain today is that WeChat has connected the file entry to Xiaowei, but complex scenarios — long PDFs, multi-sheet Excel files, PPT chart recognition — still need further testing.
A hidden capability: lightweight apps from one sentence
Xiaowei's settings also include a "Mini Tools" entry, where users describe a need in natural language to generate a simple app that runs inside the WeChat ecosystem. We typed "generate a diet recipe book for me" and it produced an interactive page with today's picks, a seven-day weekly menu, and a section for adding new recipes.
Current limitations: Mini Tools cannot yet be shared with other users, publishing is not open, and they cannot freely read WeChat context (so "automatically count the votes in this group" is not supported for now). But given that the underlying layer is the Mini Program system, the imagination space grows considerably if sharing, collaboration, distribution — even payments — open up down the road.
Who should use it, and how
Xiaowei's most valuable capabilities at this stage are summarizing group messages, recalling chat history, and understanding conversational context. For heavy WeChat users with large friend lists and cluttered chat records, and for anyone juggling many work groups with massive message volume, this is a genuine must-have. By comparison, capabilities like sending messages or making WeChat calls show no obvious advantage yet.
Xiaowei may not yet be the strongest AI assistant, but it is already the AI that comes closest to WeChat's own form — the point is not launching another assistant, but hiding AI inside every high-frequency information entry point.
Toolin Editorial Team
Categories
Related articles

Hands-On with Doubao Pro: An Agent Office Mode for 200 Million People
Doubao has launched a Pro tier with an agent-driven office task mode powered by Doubao 2.1 Pro — it can operate your local computer, analyze earnings reports, and build its own Skills; hands-on delivery quality benchmarks against Claude Opus.

Netflix Open-Sources Headroom: Cutting 90% of Wasted Tokens and Saving $700,000
Headroom, an open-source tool built by a senior Netflix engineer, trims redundant tokens before prompts ever reach the LLM — reversible compression, lossless context — and has already saved users $700,000.

Running LLMs Locally, Hands-On: Building an Agent Coding Environment with Pi + LM Studio
Use the Pi agent framework with the LM Studio inference engine to run Gemma 4 on a local Mac for agent coding tasks, with the complete Docker configuration and a models.json example included.

STAR-PólyaMath: The Open-Source Reasoning Framework That Teaches LLMs to Correct Their Mistakes
A multi-agent reasoning framework open-sourced jointly by Tsinghua and Microsoft, where the Reasoner, Verifier, and Meta-Strategist roles make long-horizon reasoning verifiable and traceable — beating GPT-5.5 on the Apex benchmark by 13.5%.

Baidu Open-Sources Unlimited OCR: Reading an Entire Book in One Forward Pass
Unlimited OCR, Baidu's model built on DeepSeek OCR, achieves 32K-context long-document parsing through the R-SWA mechanism, with an end-to-end SOTA of 93.23% on OmniDocBench v1.5.

FineVLA Goes Open Source: One Sentence Tells the Robot Which Hand to Use and Where to Grab
HKU and Alibaba jointly open-source FineVLA, a controllable VLA framework that lets language specify the executing arm, contact region, and other details, with an 86.8% success rate in RoboTwin simulation.