I’ve been quietly watching Tencent’s WorkBuddy for a few months. It’s not the flashiest AI tool out there, but the numbers are hard to ignore. Over 20 million desktop visits in June alone. That’s not nothing.
Now they’re about to drop Hy4 into it.
Here’s the short version: Tencent’s new flagship model, Hy4, was spotted in a leaked interface recently. It’s labeled expert-level, sitting above Hy3 and DeepSeek in the model picker. The official release is coming any day now. They teased it in their Q2 earnings call last week. But honestly? The leak isn’t the story. The story is what Hy4 means for WorkBuddy.
WorkBuddy Is Already Doing Something Right
Let me back up.
WorkBuddy is Tencent’s productivity agent. Think of it as an AI coworker that handles coding, report generation, frontend design, financial modeling, that kind of thing. Hy3 powers it today. And Hy3 has been genuinely solid. I’ve written about where it fits next to Codex, and the short answer is they’re not really rivals.
I looked at the numbers. 295 billion total parameters, 21 billion active. In internal blind tests with 270 experts, Hy3 scored 2.67 out of 4, beating GLM-5.1’s 2.51. That might not sound huge, but in productivity tasks like software dev, office work, and data analysis, the gap was noticeable.
And here’s the stat that made me pause: WorkBuddy’s task completion rate jumped from 72% to 90% with Hy3.
That’s a real improvement. Not a marketing number.
So What Does Hy4 Bring?
We don’t have the spec sheet yet. I checked everywhere. No official numbers on parameters, speed, or benchmark scores. But the expert-level label tells us something.
Hy4 is positioned above Hy3. That means it’s not just a small tweak. They’re aiming it at complex, multi-step problems, the kind of stuff that breaks simpler models. Think about a coding agent that needs to refactor an entire module, or a financial model that requires reasoning across messy spreadsheets. That’s where expert matters.
The rumor mill says two things could be happening under the hood.
One: they’re scaling parameters up toward a trillion, following the trend we’re seeing from other big labs.
Two: they’re doing more reinforcement learning on top of Hy3’s foundation, kind of like what Zhipu did going from GLM-5.2 to 5.3, where performance jumped significantly without a massive parameter increase.
I honestly don’t know which it is. Could be both. Could be neither. Tencent’s been tight-lipped about Hy4’s architecture. But for WorkBuddy users, the architecture matters less than the outcome.
There’s a practical angle worth mentioning. A bigger or better-tuned model only helps if the agent around it can actually use it well. WorkBuddy’s value isn’t just the raw model, it’s the wrapper that turns a prompt into a finished artifact. If Hy4 is sharper but the tooling stays the same, the gains might be smaller than the label suggests.
Why This Actually Matters
Here’s the thing about productivity AI. It doesn’t need to be the smartest model on the planet. It needs to be reliably useful.
Hy3 already proved it could handle real work. The 90% completion rate tells me that. So if Hy4 moves that needle even a little, WorkBuddy becomes genuinely hard to ignore for teams doing serious development or data work.
I’m not saying it’ll replace human workers. The point is, if your AI agent can take a messy project brief and turn it into working code or a structured report, that changes how you allocate time. You spend less time on grunt work, more time on decisions.
Plus, Tencent’s hybrid strategy is smart. They run their own models alongside open-source ones like DeepSeek. If Hy4 is genuinely stronger on complex reasoning, they’ll use it for those tasks. If DeepSeek is better at something else, they’ll route to that. It’s pragmatic. No vendor lock-in theater. Running DeepSeek alongside a flagship is cheap enough that the math actually works.
One caveat I’d throw in: more capability usually means more places for the tool to confidently do the wrong thing. A smarter model that misreads your intent can be more dangerous than a dumber one that clearly can’t.
What I’m Watching For
I haven’t used Hy4. Nobody outside Tencent has, as far as I can tell. So I can’t say it’s amazing. But I’m watching a few things:
First, WorkBuddy’s response quality on messy inputs. Can Hy4 handle vague instructions better than Hy3? That’s the real test. Most real prompts are half-baked, not the clean examples in a demo.
Second, speed. An expert model that’s slow isn’t useful in a productivity tool. You need answers in seconds, not minutes.
Third, cost. WorkBuddy runs on Tencent Cloud, so I’m curious if Hy4 drives up usage costs or if they’ve optimized it to stay affordable.
I could be wrong about all of this. Maybe Hy4 is just a minor bump. But the timing is suggestive. Hy3 landed in July, Hy4 is coming now, which tells me Tencent isn’t treating this as a once-a-year refresh.
A small tip for later: keep a fixed set of tasks you run through every model upgrade, same prompt and same files. That way you can tell whether the new version is genuinely better or you’re just feeling the novelty.
Would I Recommend WorkBuddy After Hy4?
Ask me in a month.
Seriously. Right now, Hy3 is solid. If you’re already using WorkBuddy, you’ll probably appreciate the upgrade when it rolls out. If you’re not using it, wait and see. Let the early adopters hit the bugs first. I learned that lesson the hard way with a different tool last year.
Here’s what I keep coming back to though: 20 million visits in a month isn’t hype. That’s real usage. People are getting value out of WorkBuddy today. Adding an expert-level model on top of that? It could push something already working into genuinely impressive territory.
Or it could be a nothingburger. We’ll know soon.
I’ll report back once I actually get my hands on it. For now, I’m curious. Not convinced, but curious. That’s about as honest as I can be. If you want a more skeptical take, I’d suggest watching the independent benchmarks rather than the launch announcement. The announcement will tell you what they want you to believe. The benchmarks, once they’re out, will tell you what you’ll actually get.
