WeChat launched its AI assistant, "Xiaowei". Will it become a new battlefield for large AI models?

A few days ago, many people noticed a small green eye appeared in the upper left corner of their WeChat main interface. When they clicked on it, a dialog box named "Xiaowei" popped up, with the words "Beta Version" at the top.

No one officially announced it in advance, there was no press conference, and even Tencent's customer service only confirmed it after being questioned by the media. This was the first public appearance of WeChat's native AI assistant.

According to the introduction, Xiaowei can operate WeChat's native functions through text or voice, such as adjusting settings, sending messages, making calls, and ordering takeaways. Xiaowei can also launch mini-programs to complete services like registering for medical appointments, hailing a taxi, booking a hotel, and topping up phone credit. In NetEase Technology's experience report, Xiaowei can also search for content on official accounts, generate pictures, and even "generate a mini-program with one sentence".

Also this month, Apple announced Siri AI at WWDC. Unexpectedly, the WeChat version of "Siri AI" appeared so soon. What makes people even more curious is, when "Xiaowei" can access WeChat's huge ecosystem, will it become the largest Internet portal?

Meanwhile, just as Apple's Siri AI will integrate with ChatGPT and Gemini, will WeChat's Xiaowei become an entry point for Chinese large models to compete for access?

01. WeLM and DeepSeek Behind Xiaowei

The model running behind Xiaowei is called WeLM, a self-developed product of WeChat's AI team. However, if you search for information about WeLM now, you might easily fall into a trap. The most widely circulated introduction online is actually about the "old" WeLM in 2022, first disclosed by Tencent Cloud Developer Community.

That generation of WeLM was a Chinese model with tens of billions of parameters, following the same autoregressive path as GPT-3 and Google PaLM. It had three versions with sizes of 1.3B, 2.7B, and 10B. The official training corpus announced at that time was 10TB of raw data, which was reduced to 262B tokens after cleaning. It was claimed to be able to match models 25 times larger than itself in 14 Chinese NLP tasks. The calling method was also very "academic" - there was no chat interface, you had to fill out a form to apply for a token, and then run it through the API, with a speed limit of 30 requests per minute.

This is no longer the WeLM running behind today's Xiaowei.

The WeLM team disclosed on its official technical blog that the new generation of WeLM has adopted a highly sparse MoE (Mixture of Experts) architecture, combining loss-free balance routing, an unnormalized sigmoid gate, and a shared expert. In terms of network structure, it borrowed from the Qwen series and made it deeper.

In terms of scale, the team trained an 80B - A3B MoE model on a corpus of less than 14T tokens. Its performance is said to be better than systems of the same or even larger scale. Then, through depth up-scaling, they created a 130B variant, which showed significant improvement over the basic version with only a small amount of additional training. The team mentioned in the same blog that the model number of the previous generation was WeLM-V3-258B-A22B.

From a 10B autoregressive small model to an 80B - A3B/130B MoE large model, this is not just an "upgrade", but a complete architectural replacement.

It's not just WeLM. To make WeChat's Xiaowei run better, the WeChat team has also integrated other models to ensure a good experience for Xiaowei.

02. The Battlefield of AI Large Models?

According to a report by Sohu.com, Tencent's customer service responded that Xiaowei "currently uses multiple AI models, including self-developed and high-quality open-source models by Tencent". NetEase Technology's report mentioned that according to Xiaowei's own response, the main model is Tencent's self-developed WeLM, and some answers are supported by DeepSeek.

This shows that Tencent is not betting all on its self-developed model.

A reasonable guess is that WeLM is responsible for core dialogue scenarios that require in-depth integration of WeChat ecosystem knowledge, long-term memory, and context management, while some general Q&A and knowledge-based tasks are handed over to more cost-effective open-source models like DeepSeek. In short, this is a practical division of labor of "self-developed to ensure the foundation, open-source to supplement efficiency", rather than trying to do everything on its own.

Coincidentally, Apple's choice on the other side of the ocean is exactly the same. The Siri AI unveiled at WWDC, in addition to Apple's self-developed Apple Foundation Model, has also integrated Google's Gemini model, and it is very likely to integrate more third-party models in the future.

No company dares to entrust the lifeline of a super portal entirely to a self-developed model. This might be the fact that is more worth remembering than comparing parameters and rankings in this round of AI assistant battles.

Just as Apple didn't bet all on ChatGPT but introduced Gemini, and may introduce Claude in the future, will WeChat's AI assistant Xiaowei also open its interface to other manufacturers in the future?

After all, with WeChat's huge scale, for any large model manufacturer, accessing it is a huge C-end traffic portal. Of course, for any manufacturer, accessing WeChat's Xiaowei also means facing the same challenges:

Computing power/stability: Can large model manufacturers meet the huge inference needs of WeChat's more than one billion users?

Business monetization challenge: As a super portal, Tencent usually charges the implanted parties or supports its partners to compete with other giants. It is entirely possible that large model parties may have to bear the huge inference costs from WeChat's Xiaowei.

Data privacy: Large model manufacturers need data for training, but WeChat clearly attaches great importance to data privacy. How user conversations and service data will flow also depends on WeChat's "attitude".

However, it cannot be denied that during the upgrade process of WeChat and Tencent's self-developed models, integrating third-party models to ensure a good experience still provides a perfect C-end training scenario for all large model companies.

Looking back, Xiaowei's real competitor may be Doubao from ByteDance, currently the largest AI application in China. With more than 300 million users, Doubao is in the first echelon of domestic AI applications. WeChat's Xiaowei, on the other hand, may use the advantages of the WeChat ecosystem to gain an edge in the AI portal battle for Tencent.

03. The Battle for the Portal in the AI Era

The really interesting thing about Xiaowei is that it and Siri AI are actually two different ways to achieve the same ambition.

Siri AI has a clear goal: to become the unified AI portal for the entire iOS operating system. Craig Federighi, Apple's senior vice president of software engineering, said at WWDC that the new version of Siri can access and operate all applications and data on users' iPhones. It supports reading emails, calendars, and screen content, and can perform cross-application tasks. This is a system-level ambition.

Obviously, Xiaowei can't reach this level. It doesn't have system permissions and can't do anything outside the WeChat app.

But on the other hand, WeChat itself is a "system" in China - the chat relationship chain, more than one billion monthly active users, the service network composed of millions of mini-programs, the content accumulated on official accounts and video accounts, and the consumption scenarios bound to WeChat Pay. The combination of these resources is almost the scale of another operating system, except that it runs on top of iOS and Android, rather than on the underlying hardware.

So what Xiaowei does is essentially to put the "system-level ambition" into an "App-level container". It doesn't need to obtain the permissions from mobile phone manufacturers. It only needs to connect all resources into a call chain within the WeChat boundary. With just one sentence from you, it can help you order coffee, check express delivery, book a hotel, and generate a mini-program.

This is a " unified portal within the ecosystem", rather than a real system unified portal. This qualifier is very important - it means that Xiaowei's ceiling is essentially the ceiling of the WeChat app. It won't be higher than that, but because of WeChat's large scale, this ceiling is already high enough.

Just before the release of Xiaowei, WeChat Pay launched an "AI Exclusive Card" - a card designed specifically for AI agent payment scenarios and built into the change. It has currently been integrated with Tencent's desktop office efficiency AI agent WorkBuddy, which can complete the whole process from recommendation to placing an order and making a payment. Xiaowei is only one step away from a "closed loop": to integrate decision-making, execution, and payment all within WeChat, so that users don't need to leave the app.

Tencent won't let this opportunity slip away. With more than one billion monthly active users, whoever turns the "AI portal" into an "ecosystem closed loop" first will get the ticket to the next decade of the mobile Internet.

Xiaowei is still rough and in the gray testing stage, but it may have more time and a larger sample size for iteration than any competitor.

*The source of the header image: AI-generated

This article is from the WeChat official account "GeekPark" (ID: geekpark), written by Hualin Wuwang, and published by 36Kr with authorization.