46
0Episode: GLM-5.3 Flash: Same Brain, 1/40 the Price
Duration: approximately 9 minutes
Level: B1 (Intermediate)
---
[Mike]: Welcome back to Learn English with Podcasts! Sarah, quick riddle - what costs 40 times less but thinks just as well?
zh:欢迎回到 Learn English with Podcasts！Sarah，猜个谜语——什么东西便宜 40 倍，脑子却一样好使？
[Sarah]: Is this a sale at the supermarket? Or a new phone plan?
zh:是超市大促销？还是新的手机套餐？
[Mike]: Better. It is a brand-new AI from China. Zhipu AI just released and open-sourced GLM-5.3 Flash. 320B total size, but only 18B active at a time. And its brain is frontier-level.
zh:更厉害。是一颗来自中国的全新 AI。智谱刚发布并开源了 GLM-5.3 Flash。总参数 320B，但每次只激活 18B。脑子却是前沿水平。
[Sarah]: 320B-A18B - that sounds like a giant library where you only open one shelf at a time.
zh:320B-A18B——听起来像一个巨大图书馆，但每次只打开其中一排书架。
[Mike]: Perfect picture. And here is the number that made everyone stop scrolling. On the Artificial Analysis Intelligence Index, a very respected global test, it scored 57.
zh:这个比喻完美。而让所有人停下刷手机的数字是这个：在非常权威的全球测试 Artificial Analysis Intelligence Index 上，它得了 57 分。
[Sarah]: 57 means... good? Great? Help me feel it.
zh:57 分算……好？非常好？让我感受一下。
[Mike]: 57 is frontier. Exactly the same as Anthropic's most popular model, Claude Opus 4.8. On Z.ai's own Code Bench, which tests real coding feel, it also ties with Opus 4.8.
zh:57 就是前沿。和 Anthropic 最受欢迎的模型 Claude Opus 4.8 完全同分。在 Z.ai 自己的 Code Bench 体感测试里，编程表现也和 Opus 4.8 打平。
[Sarah]: Same brain, then. So what is the price part of your riddle?
zh:那就是脑子一样。所以你谜语里的价格部分呢？
[Mike]: GLM-5.3 Flash costs 1/10 of the full GLM-5.3. With the current limited-time discount, it is 1/20. And compared to Opus 4.8, it is 1/40.
zh:GLM-5.3 Flash 的定价是完整版 GLM-5.3 的 1/10。现在限时折扣内是 1/20。和 Opus 4.8 比，是 1/40。
[Sarah]: Wait, let me do the math. Same 57 points, 1/40 the price? That is not a discount, that is a different universe.
zh:等等，我算一下。同样的 57 分，1/40 的价格？这不是打折，这是换了个宇宙。
[Mike]: Their slogan says it: same intelligence, 1/40 the price. Frontier AI, finally cheap enough that you do not have to save it.
zh:他们的口号就是这么说的：同样智力，1/40 价格。前沿智能，第一次不用省着用。
[Sarah]: Okay, I love a good slogan, but how do we know it is real? Companies can claim any number.
zh:口号我喜欢，但怎么知道是真的？公司可以随便说数字。
[Mike]: They did a blind test. Before the launch, they released it anonymously as Ox-Alpha - in Chinese, the community called it Niu Lai. On OpenCode and OpenRouter, no name, no logo.
zh:他们做了盲测。发布前，他们匿名把它放出来，代号 Ox-Alpha——中文社区叫它牛来。在 OpenCode 和 OpenRouter 上，无名字无 Logo。
[Sarah]: And people just... started using it without knowing what it was?
zh:然后大家就……在不知道它是谁的情况下开始用？
[Mike]: And loved it. It quickly became the most popular model of the week and broke call records on both platforms. Only later they said, surprise, that was us.
zh:而且很喜欢。它很快成了当周最受欢迎的模型，创下双平台调用量纪录。后来他们才说，惊喜一下，那就是我们。
[Sarah]: Okay, that is confident. But you said same frontier brain for less money. Where does the savings come from? Smaller model?
zh:好，这很自信。但你说同样的前沿脑子花更少的钱。省钱从哪来？把模型变小？
[Mike]: The opposite trick. Total size is similar to the older GLM-4.5 - 320B vs 355B. But look inside. Active parameters dropped from 32B to 18B. Layers nearly halved, from 92 to 45. All trained on a new 30T token multimodal dataset.
zh:恰恰相反的技巧。总参数和上一代 GLM-4.5 差不多——320B 对 355B。但看里面。激活参数从 32B 降到 18B。层数几乎减半，从 92 层到 45 层。全部用全新的 30T token 多模态数据训练。
[Sarah]: Fewer active parts, fewer layers, but stronger result? That sounds like losing weight and getting stronger at the same time.
zh:激活更少、层数更少，结果却更强？这像减重同时变壮，一起发生。
[Mike]: Exactly. And for long contexts, they built a new attention system. It mixes two ideas: linear attention for nearby words, and sparse attention for far-away context.
zh:没错。而为了处理长上下文，他们造了一套新的注意力系统。混合两种思路：线性注意力管附近的词，稀疏注意力管远处的上下文。
[Sarah]: Like having both a flashlight for the page in front of you, and a search light for the whole library?
zh:就像既有照亮眼前这一页的手电，又有扫完整座图书馆的探照灯？
[Mike]: Beautiful. To save even more, they added IndexPool. It squeezes 4 memory vectors into 1 for the 1M context index. Much less memory, much less delay.
zh:太形象了。为了再省一点，他们加了 IndexPool。它把 1M 上下文索引里的 4 个缓存向量压成 1 个。内存和延迟都大幅下降。
[Sarah]: Numbers please. How much did they actually save?
zh:给我数字。到底省了多少？
[Mike]: Compared to the full GLM-5.3, attention compute is down 3.01 times, and KV cache size is down 4.44 times per layer. Among all the models they compared, including DeepSeek-V4-Flash and Kimi-K3, this one has the lowest compute per token.
zh:和完整版 GLM-5.3 比，每层注意力计算量降了 3.01 倍，KV 缓存大小降了 4.44 倍。在他们对比的所有模型里，包括 DeepSeek-V4-Flash 和 Kimi-K3，它的单 token 计算量最低。
[Sarah]: But you said lowest compute, so what is still not perfect?
zh:但你说计算量最低，那还有哪里不完美？
[Mike]: Honest detail: its KV cache is still a bit bigger than Kimi-K3 and DeepSeek-V4-Flash. The team says that is their next job.
zh:诚实的细节：它的 KV 缓存仍然比 Kimi-K3 和 DeepSeek-V4-Flash 略大。团队说那是下一步要优化的。
[Sarah]: I like that honesty. Now, you said this is the first native multimodal in the GLM-5 family. What does multimodal mean here? It can see?
zh:我喜欢这种坦诚。对了，你说这是 GLM-5 系列首个原生多模态。说的多模态在这里是什么意思？它会看？
[Mike]: It can see, and it uses seeing to code better. They call it Visual Coding. For frontend, games, 3D, the final result is something you see and touch. So the model learns when to look at its own output and fix it.
zh:它会看，而且用看来看得更会写代码。他们叫视觉编码。做前端、游戏、3D 时，最终产物是你能看到、能交互的东西。所以模型学会何时去看自己的输出并修正。
[Sarah]: So it is not just write code, run, hope it works. It is write, look, fix, like a human designer?
zh:所以不是写完代码跑一下就祈祷成功。而是写、看、改，像人类设计师那样？
[Mike]: Exactly. They built a data pipeline where the model has to interact, check its own screen, and improve. With reinforcement learning from real user flows, it even judges if a button looks right, not just if the code runs.
zh:没错。他们搭了一条数据流水线，让模型必须去交互、检视自己的画面、再迭代改进。加上基于真实用户流程的强化学习，它甚至能判断按钮好不好看，而不只是代码能不能跑。
[Sarah]: Give me a wild example. What did it actually build?
zh:给我一个疯狂的例子。它到底做出了什么？
[Mike]: My favorite. With no outside images or models, GLM-5.3 Flash ran alone for 16 hours in Blender and built a 400-square-meter professional chef's home and test kitchen. Every piece of furniture, light, and material placed consistently from any angle.
zh:我最喜欢这个。在没有任何外部素材的情况下，GLM-5.3 Flash 在 Blender 里独自跑了 16 个小时，搭出了一套约 400 平方米的专业主厨自宅和测试厨房。每一件家具、每一束光、每一种材质，从任何角度看都一致。
[Sarah]: 400 square meters, 16 hours, no help? That is my apartment times four, built while I slept.
zh:400 平方米，16 小时，无协助？那是我家四倍大，在我睡觉时就盖好了。
[Mike]: And inside ZCode, it can now work across code, browser, and computer screen together. Many bugs only show up after you render or click around.
zh:而且在 ZCode 里，它现在能在代码、浏览器和电脑界面之间协同工作。很多问题只有渲染出来、点一点才暴露。
[Sarah]: Okay, coding is impressive. But most people at work just want a good PPT, not a 3D kitchen.
zh:编程很厉害，但大多数上班族只想要一份好看的 PPT，而不是 3D 厨房。
[Mike]: It does that too. With visual understanding, it can compare its own PPTX, PDF, DOCX, and XLSX output to what it expected, and fix the beauty. Plus they trained it for finance and law.
zh:那个也行。靠视觉理解，它能把自己输出的 PPTX、PDF、DOCX 和 XLSX 与预期对比并优化美观。还专门为金融和法律做了训练。
[Sarah]: Finance and law? Like real reports?
zh:金融和法律？像真正的报告那种？
[Mike]: Yes. In finance: from research with sources to report writing to modeling, all with traceable references. In law: it can review cost and liability clauses, mark them like a real lawyer, and draft letters and contracts ready to send.
zh:对。金融那边：从带来源的研究、到报告生成、再到建模分析，全流程可追溯。法律这边：能审费用和责任条款，按律师习惯批注留痕，还能起草函件和合同，格式规范到可直接交付。
[Sarah]: Ready to send? No more my 2 a.m. formatting nightmare?
zh:可直接交付？再也不用凌晨两点调格式了？
[Mike]: That is the promise. Now the last twist, and it is a big one. Remember Ox-Alpha broke records? All that traffic ran on domestic Chinese chips.
zh:这就是承诺。还有最后一个大反转，也很大。还记得牛来创下纪录吗？那些流量全部跑在国产芯片上。
[Sarah]: Wait, all those calls were served by Chinese chips? Not the usual big brand?
zh:等等，那些调用全是国产芯片扛住的？不是常用的大品牌？
[Mike]: A full domestic cluster, connected by their own high-speed network. Single chips have less memory and bandwidth, especially for 1M context. So they built a special engine on SGLang.
zh:一整套国产芯片集群，用自研高速网络连起来。单张芯片内存和带宽相对有限，特别是要撑 1M 上下文。所以他们基于 SGLang 造了专用推理引擎。
[Sarah]: And who built that engine? A huge team working for months?
zh:谁造的这个引擎？一支大团队干了几个月？
[Mike]: Here is the fun circle. A big part of the engine was built faster with help from an infra agent powered by GLM-5.3 itself. The model helped optimize the system that now runs the model.
zh:好玩的闭环来了。引擎很大一部分是靠 GLM-5.3 自己驱动的 infra agent 加速搭起来的。模型帮忙优化了现在承载模型的系统。
[Sarah]: The model improved the system that carries the model. Like a student who builds a better desk to study faster?
zh:模型改进了承载自己的系统。像学生为了学得更快先造了一张更好的书桌？
[Mike]: Exactly. They used tricks like intra-node tensor parallel for linear attention, ReplaySSM, W8A8 quantization, mixed INT8/FP8/BF16 cache, and Layer Split. At the cluster level, they split Encode, Prefill, and Decode into separate pools that scale alone.
zh:没错。他们用了很多技巧：线性注意力的节点内张量并行、ReplaySSM、W8A8 量化、INT8/FP8/BF16 混合缓存、Layer Split。集群层面还把编码、预填充和解码拆成独立可扩缩的工作池。
[Sarah]: Okay, translate that to human: did it work?
zh:好，翻译成人话：管用吗？
[Mike]: Compared to the first baseline on the same hardware, end-to-end speed went up 3 times. And cost per token is now similar to mainstream NVIDIA GPUs.
zh:和同一硬件上的初始基线比，端到端性能提升了 3 倍。单 token 成本已和主流英伟达 GPU 相当。
[Sarah]: So domestic chips can now do frontier work at a similar price. That changes the story from who has the best chips to who uses them best.
zh:所以国产芯片现在也能以差不多的价格扛前沿任务。故事就从谁有最好的芯片，变成了谁把芯片用得最好。
[Mike]: And it is open now. API on BigModel and Z.ai, chat on chat.z.ai, plus ZCode and AutoClaw. With GLM Coding Plan giving 10,000 free cards a day.
zh:而且现在已开源开放。BigModel 和 Z.ai 上可调 API，chat.z.ai 可直接体验，还有 ZCode 和 AutoClaw。GLM Coding Plan 每天还发 10,000 张体验卡。
[Sarah]: You know what sticks with me? Not the 57, not the 1/40. It is the 16-hour kitchen. A model that can stare at its own work, say hmm, that light looks wrong, and fix it - that feels... human.
zh:你知道最让我记住的是什么吗？不是 57 分，也不是 1/40。是那间搭了 16 小时的厨房。一个会盯着自己的作品说嗯这灯光不对然后改掉的模型——感觉……很人类。
[Mike]: And the quiet joke behind it all - weeks ago, thousands of people fell in love with a nameless cow. Niu Lai. They had no idea they were already living in the cheaper future.
zh:而背后那个安静的玩笑是——几周前，成千上万人爱上了一头无名牛。牛来。他们完全没意识到，自己已经活在了更便宜的未来里。
[Sarah]: Same brain, tiny bill, and a cow that tricked everyone. I think our listeners will remember that. Thanks for listening to Learn English with Podcasts - see you next time!
zh:同样的脑子，更小的账单，还有一头骗过所有人的牛。我想听众会记住这个。感谢收听 Learn English with Podcasts——下次见！
[Mike]: Try GLM-5.3 Flash and tell us what you would build - a kitchen, a deck, or your dream app. See you next time!
zh:试试 GLM-5.3 Flash，告诉我们你会做什么——一间厨房、一份演示文稿，还是你梦想中的 App。下期见！
