Original video in Chinese.
Key Takeaway
- Tencent Yuanbao is the best introductory AI tool for ordinary users, thanks to its all-scenario coverage, dual-model support, and Tencent-enhanced R1.
- Yuanbao offers a web version, desktop clients, a mobile app, and a WeChat mini program, and supports importing files from WeChat, enabling seamless switching across scenarios.
- Yuanbao combines DeepSeek-R1 (slow thinking) and Hunyuan Turbo S (fast thinking) in a dual-model setup, balancing depth and efficiency.
- Tencent has enhanced R1, including access to public account article search sources and image recognition, making it a “full-power” R1.
- The article emphasizes Tencent’s strong product capabilities and in-house development capabilities in the consumer market, as well as the significance of AI popularization for ordinary users.
The DeepSeek craze has definitely made Tencent the happiest company.
Because when every company integrates DeepSeek and the models are all the same, once model capabilities are leveled, what are they really competing on? Product. And when it comes to product, who has Tencent ever been afraid of? Not to mention that, besides DeepSeek, Tencent also has its in-house Hunyuan large language model.
Have you noticed that during this period after the New Year, Tencent, which has always been very Buddhist and very casual, suddenly became much fiercer and more decisive?
On the B-side, Tencent Cloud has integrated DeepSeek into its core product line, and has also launched cloud services and solutions based on DeepSeek.
On the consumer side, Yuanbao shot up to second place on the free chart in the App Store, beating Kimi and Doubao, and coming in just behind the official DeepSeek app.
I think that for ordinary people, starting to use AI with Yuanbao is definitely the best AI introduction tool. Most of the videos I’ve posted before are aimed at users with some foundation. Then every one or two days, someone privately messages me asking: how should a complete beginner get started with AI?
It’s simple: in the AI era, the most efficient way to learn is Learning by doing. Actually getting hands-on and using it frequently in daily life matters more than reading a hundred tutorials. And Yuanbao is the first entry point for ordinary users who are not early adopters. Three reasons.
First, all-scenario coverage.
Using AI is actually the same principle as using a camera. The camera you can carry with you is the good camera. Likewise, the product that lets you use AI anytime is the good product.
Yuanbao has a full-platform web version, desktop clients for both Mac and Windows, mobile apps for both iOS and Android, and a WeChat mini program that requires no installation.
Its real-time synchronization across all four endpoints perfectly adapts to switching between scenarios. For example, a conversation that you didn’t finish on the desktop can be continued in the app when you’re out and about. I usually pin the chats that need repeated discussion, so next time I don’t have to dig through my history for ages.
On top of these basic features, Tencent did something only Tencent can do:
Import documents from WeChat.
I believe more than 90% of users treat WeChat’s “File Transfer Assistant” as a “file storage assistant” and dump everything into it. Because on mobile, WeChat is the first entry point. It makes perfect sense to look for documents inside WeChat.
So naturally, on a phone, if you want to hand a document to AI for analysis, importing it directly from WeChat is the most convenient option.
So when you tap the plus button in the bottom-right corner of the Yuanbao app, you’ll see the “WeChat Files” option. It jumps to WeChat and then uses the Yuanbao mini program as a bridge to make WeChat file transfer work. Or, conversely, you can directly choose Yuanbao by opening it in “other apps” inside WeChat.
See, this is what product capability means. It looks like a tiny feature, but it’s practical and thoughtful. This is a textbook Tencent style feature.
I really hope there will be more integration between WeChat and Yuanbao going forward. For example, on a very long public account article, when you tap the share button in the top-right corner, there could be an “Open in Yuanbao” option. Then, just like the current “deep reading document” feature, it could offer three options: summary, in-depth reading, and mind map. I’m sure this would definitely become a hit feature.
Second, dual-model support.
DeepSeek-R1 is great, but it’s not suitable for every use case.
Models like R1 are reasoning models. Through multi-stage training, especially reinforcement learning, they ultimately gain the ability to reason step by step and build chain-of-thought. So when they face complex problems, they first analyze the problem, identify key points, and then gradually arrive at conclusions.
But in our daily use, are all questions complex questions? Are all questions worth waiting several seconds for an answer?
Obviously not. So just having DeepSeek-R1 is not enough; you need other models too. This is where the importance of in-house development capabilities shows up.
A couple of days ago, Tencent launched the Hunyuan Turbo S model. If DeepSeek-R1 and Hunyuan T1 are called “slow thinking” models, then Hunyuan Turbo S is a “fast thinking” model. It has lower time to first token and faster output speed. In two words:
Instant reply.
In Yuanbao, when you choose the Hunyuan model and turn off the “deep thinking” button, you’re using Hunyuan Turbo S for quick Q&A.
Honestly, I used R1 for a period before, and every time I saw a simple question and it was over there mulling over it for ages, I was extremely speechless. For example, if I said “hello,” it would think for 5 seconds before replying. Its inner monologue was way too strong, with a kind of cyberpunk absurdity.
So, “slow thinking” plus “fast thinking,” with a dual-model combination, is what makes daily use efficient.
Third, Tencent-enhanced R1.
Modern large language models have evolved into “application platforms,”
Two years ago, models were basically just text chat. But today’s models have many more key functions. For example, look at Google’s Gemini 2.0, which emphasizes the concept of “native” — native image generation, native text-to-speech, native tool use.
The goal of this big integration is to build a universal AI assistant. This has also become standard for flagship models now.
Mainland media, in the pursuit of traffic, when praising DeepSeek-R1, only focus on its reasoning ability and deliberately leave out the rest. So the task of adding more tools and more functions to R1 falls to application vendors. In this regard, Yuanbao has two things I especially like.
First, the search sources include public account articles.
The Chinese internet has gotten full of sand and mud over the past few years, and there isn’t much content worth reading left. So when I use tools like Perplexity, I specifically emphasize in the prompt: even if I ask in Chinese, you should search English web pages for me.
If there is still anywhere with decent content, then it’s probably only public accounts left. The content on this platform is Tencent’s exclusive resource. Yuanbao’s search now makes use of that.
High-quality public account content, combined with DeepSeek-R1’s strong reasoning ability, equals quality in, quality out.
Second, image recognition capability.
R1 itself does not have image recognition capability. So applications, including the official DeepSeek app, add an OCR function to extract the text from images and then pass it to the model for further processing.
But if there’s no text in the image and I just want to analyze a pure image?
That is Tencent’s second enhancement to R1. They first analyze the image content and then give the result to R1 as a reference for answering.
For example, I took a photo of my desk setup. On the DeepSeek website, it will prompt “No text extracted.” But in Yuanbao, R1 can accurately describe the image.
Running an R1 with 671B parameters doesn’t make it a full-power version. Being able to search good content and read images is what a full-power R1 really is.
The emergence of DeepSeek-R1 is a very clear signal to the industry. Many companies received that signal and took action, but only Tencent embraced it most thoroughly.
When you have massive traffic and extremely strong product capabilities, making a comeback in the consumer market and acquiring users is just a matter of a shake — internal traffic diversion, external advertising, who can’t do that? When you have sufficiently strong in-house development capabilities and can build differentiated models, then you won’t get tangled up in whether to integrate open-source models.
So once Tencent enters the battlefield, all the companies in the industry, big and small, have to think about what to do next. And for ordinary users, get on board quickly — the barrier to entry is already very low. As I always say:
AI won’t replace you; people who use AI will.
OK, that’s all for this episode. If you want to learn about AI, come join our newtype community. See you next time!