Hunyuan Hy4 Preview: Open-Source Top Tier
Welcome back to Learn English with Podcasts! Sarah, imagine an AI that could do almost your whole workday - write code, check your invoices, even build a small video game - and this week you can try it for free.欢迎回到 Learn English with Podcasts！Sarah，想象一个 AI 几乎能搞定你一整天的工作——写代码、核对发票、甚至做个小游戏——而且这周还能免费试用。
Free? That sounds too good. Which AI are we talking about?免费？听着太好了吧。我们说的是哪个 AI？
Tencent just released and open-sourced a new model called Hunyuan Hy4 preview. And they say it is now among the very best open-source models in the world.腾讯刚发布并开源了一个新模型，叫混元 Hy4 preview。而且他们说，它现在已经是全球最好的开源模型之一。
Wait, Tencent? The company behind WeChat? They made a top AI too?等等，腾讯？微信那家公司？他们也做出了顶尖 AI？
Yes. And here is the first big number: the model has 770 billion total parameters, but only 49 billion are active each time you use it.是的。第一个大数字是：模型总参数 7700 亿，但你每次使用时只有 490 亿被激活。
Only 49 of 770? So it is like a huge company where most people are on standby, and a small team does the actual job?770 亿里只有 49 亿？那它像一家巨型企业，大部分人待命，只有一小队人真正干活？
Perfect analogy. That design is called Mixture of Experts. You keep a giant library of knowledge, but only wake up the parts you need. It saves cost and speed.这个比喻太贴切了。这种设计叫专家混合。你保留一个巨大的知识库，但只唤醒需要的那部分。省成本又提速。
Smart. And what else is impressive?聪明。还有什么厉害的？
The memory. It can read up to 1 million tokens in one go. That is roughly a whole book, or your entire year of chat logs.它的记忆力。一次能读多达 100 万 Token。差不多是一整本书，或者你一整年的聊天记录。
A whole book at once? My notebook would be full after two pages.一次读完一整本书？我的笔记本两页就满了。
And because it is open-source, developers can download the weights, study how it works, and run it on their own servers.而且因为它开源，开发者可以下载权重、研究它的原理，还能部署在自己的服务器上。
So it is not locked inside one app. The whole world can build on it?所以它不被锁在某个 App 里。全世界都能在它基础上再创造？
That is what open-source means. Tencent shared it on Hugging Face, GitHub, and ModelScope.这就是开源的含义。腾讯把它放到了 Hugging Face、GitHub 和 ModelScope 上。
Okay, but is it actually good, or just open?好，但它真有本事，还是只是开源而已？
Tencent built it for work, not small talk. They tested it in a blind test with 163 internal experts across 203 engineering tasks.腾讯是冲着干活造它的，不是闲聊。他们做了盲测，请了 163 位内部专家，覆盖 203 个工程任务。
Blind test meaning the experts did not know which model they were grading?盲测就是说专家不知道自己在评哪个模型？
Exactly. And Hy4 preview scored 2.99 out of 4, just above GLM 5.3 at 2.92 and Kimi K3 at 2.94.没错。Hy4 preview 拿到 4 分里的 2.99，略高于 GLM 5.3 的 2.92 和 Kimi K3 的 2.94。
So it barely edged them out. Close race at the top.所以它只是险胜。顶尖这场很胶着。
Very close. Now for real examples. In office work, it read 72 invoices and three company rule books, then decided which reimbursements were allowed.非常接近。说点实际例子。办公场景里，它读了 72 张发票和三本公司规章，然后判断哪些报销能过。
Three rule books? That alone would take me an afternoon.三本规章？光这个我就得耗一个下午。
And in game development, it connected to the Unreal 5 engine through something called MCP, and built a playable shooting game just by talking to it.游戏开发里，它通过一个叫 MCP 的东西接入了 Unreal 5 引擎，光靠对话就做出了一款能玩射击游戏。
A whole game from conversation? No coding by hand?整款游戏靠对话生成？不用手写代码？
That is the claim. The developer just kept chatting to improve it. And in science, it ran a molecular simulation of 32,512 atoms at 54.9 milliseconds per step.这是它的说法。开发者只要不断对话来完善。科学方面，它跑了一个 32,512 个原子的分子模拟，每步 54.9 毫秒。
Each step takes just a few milliseconds? That sounds fast for something so small.每步只要几毫秒？那么小的体系还这么快？
They say it is 2.0 times faster than before, and one top GPU could hold 300,000 atoms. Useful for drug and material research.他们说比以前快 2.0 倍，一张高端 GPU 能装下 30 万个原子。对药物和材料研究很有用。
Okay, but here is the part that gave me chills. This model helped improve itself?好，但最让我起鸡皮疙瘩的是这点。这个模型还帮着自己改进自己？
Yes. Hy4 preview joined its own development - it suggested training methods, data plans, even optimized the computer code that runs it. End-to-end speed went up 31.8 percent.是的。Hy4 preview 参与了自己的研发——它提议训练方法、数据方案，甚至优化了运行它的底层代码。端到端速度提升了 31.8%。
A model that tunes its own engine. That is like a car that redesigns its own engine while driving.一个会调自己引擎的模型。就像一辆车一边开一边重新设计自己的引擎。
And Tencent has been rebuilding its AI infrastructure since February, shipping a big update about every two months. They release a preview first, then a final version.而且腾讯从 2 月起重建了 AI 基础设施，大约每两个月发一个大版本。他们先发预览版，再发正式版。
So the preview is like a beta you get to use early?所以预览版就像你提前用上的测试版？
Exactly. And to collect feedback, WorkBuddy and CodeBuddy are free to try for two weeks. If you want, you can build something with it today.没错。为了收集反馈，WorkBuddy 和 CodeBuddy 限时两周免费。想试的话，今天就能用它做个东西。
I like that. Listeners, if you tried Hy4 preview, what would you build first - a game, a report, or your own little assistant?这个我喜欢。听众们，如果你试了 Hy4 preview，会先做什么——游戏、报告，还是你自己的小助手？
Thanks for listening to Learn English with Podcasts. The next time you hear a model is big, ask not how big, but how much of it actually shows up to work.感谢收听 Learn English with Podcasts。下次听说某个模型很大时，别问它多大，问问它到底有多少真的来上班了。
See you next time!下次见！