I Watched Tencent Demo Work Buddy on Claude Sonnet
They have their own large language model, and a work tool that can run on more than one model. The presenter ran the demo on Claude Sonnet anyway.
TL;DR: Thursday evening in Shenzhen I visited Tencent. They have their own large language model, and a work tool called Work Buddy that lets you choose which model powers it. When they presented it to us, the presenter ran the demo on Claude Sonnet anyway. It even showed that at the bottom of the screen.
I visited Tencent, one of the large technology companies in China, on the evening of November 5 in Shenzhen, as part of an executive education program, and saw how they were thinking about AI.
What I was shown
They have their own AI model, after all — their own large language model — and they have a product called Work Buddy, which I’d never used.
Work Buddy is something like Claude Code or Claude Cowork. It also lets you choose which LLM you want to use to power it.
When they made a presentation for us, instead of using their own LLM, the presenter was actually using Claude Sonnet. It even showed that at the bottom. Someone in our group asked him about it.
Despite the fact they have their own LLM technology, they were using American technology.
I’m not going to over-read one demo. A presenter may pick the model he knows best, or the one that behaves well in a short meeting, or the one that was already open. I can only tell you what we were shown. We were shown a Tencent work tool running on Claude Sonnet.
Taking a step back
I walked out thinking about two races that get bundled together too often.
On foundation models, they were using American technology. That’s a tell, or at least a data point. Tencent can build its own model and still reach for Claude when it wants the tool to work in the room.
On robots, and on the data that teaches robots to act in the physical world, the picture is different. I’m writing that up separately. The short version is: I do not think the US is ahead in the same way.
My thought is not “China can’t build models.” They can, and they do. My thought is that, in this room, on this product, American model quality was still the thing they reached for.
So what for operators and investors
- If you buy software from Chinese platforms, don’t assume the wrapper and the model are the same thing. Work Buddy, as it was shown to us, is a tool that can sit on more than one LLM. Ask which model is actually running, in production, not just in the brochure.
- If you’re investing in the model race, one demo is not a market share number. It is a reminder that even a company with its own model may still prefer someone else’s when it has a choice.
- If you’re watching China versus the US, split the stack. Language models are one race. Robot hardware, robot data, and the jobs already being automated are another.
What would change my mind: evidence that, in Tencent’s own products used at scale inside China, their model has already displaced the American ones, and this demo was just a convenience. I don’t have that evidence. I only have what we were shown.
Open question for readers: when you watch a company demo its own AI product, do you look at the logo on the slide, or at which model name is sitting at the bottom of the screen?
Disclosure: This was an executive education program trip. No company reviewed this piece. I hold no positions in any company mentioned. My own consumer brands are manufactured in China.
Sources & notes
- First-hand, from my trip notes: visit to Tencent in Shenzhen, evening of Thursday, November 5, 2026. Work Buddy, the model picker, and the presenter using Claude Sonnet are what I was shown, as I remember them.