DeepSeek V4 Flash Vision Exp: AI Learns to See
Welcome back to Learn English with Podcasts! Sarah, quick question - can your AI read your vacation photos and actually do something useful with them?欢迎回到 Learn English with Podcasts！Sarah，问你个小问题——你的 AI 能看懂你的度假照片，然后真的做点有用的事吗？
Honestly, most AIs just say nice photo! But you're smiling like you found one that does more.老实说，大多数 AI 只会说"照片不错！"但你笑得这么开心，肯定是发现了更厉害的。
I did. DeepSeek just released something called DeepSeek V4 Flash Vision Exp. It's experimental - Exp means it's a test version.我确实发现了。DeepSeek 刚刚发布了一个叫 DeepSeek V4 Flash Vision Exp 的东西。它是实验性的——Exp 就是测试版的意思。
DeepSeek V4 Flash Vision Exp - that's a long name! What makes it special?DeepSeek V4 Flash Vision Exp——名字好长啊！它特别在哪里？
Two words: it can see. The normal V4 Flash is great with text. This new one keeps all that text power, but adds eyes.两个字：它能看见。普通的 V4 Flash 就已经很擅长处理文字了。这个新版本保留了全部文字能力，还加了一双眼睛。
Wait, so it didn't lose any text skills? Usually when you add vision, the text part gets weaker.等等，所以它的文字能力一点没丢？通常加了视觉能力后，文字部分就会变弱。
Exactly, that's the balance. In pure text tasks - like solving problems or answering knowledge questions - it scores the same as V4 Flash.没错，这就是它的平衡。在纯文字任务上——比如解决问题或回答知识类问题——它的得分和 V4 Flash 一样。
Okay, that's good. But what about the vision part? How much better is it?好，这很不错。那视觉部分呢？提升了多少？
Big jump. On agent benchmarks that need vision, the normal Flash was not strong. This Vision Exp version jumps up to almost the level of Opus-4.8.提升很大。在那些需要视觉的智能体测评里，普通的 Flash 并不强。这个 Vision Exp 版本一跃提升到几乎和 Opus-4.8 同等水平。
Opus-4.8? That's one of the top models. So this experimental model is already playing in the big league?Opus-4.8？那可是顶级模型之一。所以这个实验性模型已经能和顶尖选手同场竞技了？
Yes. And the cool part is how it works inside agent tools. You give it a big goal, and it uses vision to get there.是的。更酷的是它在智能体工具里的工作方式。你给它一个大目标，它就用视觉能力去实现。
Give me a real example. I learn better with stories.给我一个真实的例子。我听故事学得更快。
Okay, first one - a luxury Tibet travel PPT. The prompt was crazy detailed: one month, self-driving, south plus north Tibet, wild and raw style, not touristy photos.好，第一个——一份高端西藏旅行 PPT。提示词非常详细：一个月，自驾，藏南加藏北，要野性原始的风格，不要游客照。
A month in Tibet? And the AI has to make slides for rich customers?在西藏待一个月？而且 AI 要给高净值客户做幻灯片？
Yes, for high-end private tours. And get this - the final PPT needed three real pricing plans, with real photo style images, rough and powerful look. No cute filters.对，是给高端私人定制游的。而且关键是——最后的 PPT 需要三种真实的报价方案，配上真实摄影风格的图片，粗粝有力量感。不要小清新的滤镜。
And the AI did it? Just from that long text description?然后 AI 做出来了？就凭那段长长的文字描述？
With the vision ability, yes. It found the right images, matched the wild tone, and built a real-looking deck that you could actually send to customers.有了视觉能力，是的。它找到了合适的图片，匹配了野性的基调，做出了一份看起来真的能发给客户的演示文稿。
Wow, so it understands both the words and the pictures fit the feeling. What's the second example?哇，所以它既理解文字，也能让图片贴合感觉。第二个例子是什么？
This one is fun. They asked it to rebuild a developer website - the DeepSeek Harness site - with a future style. Dark blue sea, glass UI, ASCII art.这个很有趣。他们让它重做一个开发者网站——DeepSeek Harness 的官网——要未来感风格。深蓝色海洋、玻璃质感界面、ASCII 艺术字。
ASCII art? Like old computer letters making pictures?ASCII 艺术？就是用老式电脑字母拼成图片那种？
Exactly. And after many rounds of talking, the model turned that idea into a whole new site. It kept listening, kept seeing what it made, and fixed it step by step.没错。而且经过多轮对话后，模型把那个想法变成了整个新网站。它一边听，一边看自己做出来的东西，然后一步步修改。
So it's not one-shot. It's like working with a designer who shows you drafts?所以不是一次成型。就像和一个会给你看草稿的设计师合作？
Perfect example. And third one - a mini website with cute 3D clay monsters dancing at a retro party.完美的比喻。第三个——一个迷你网站，一群可爱的 3D 黏土小怪物在复古舞池里跳舞。
Clay monsters dancing? That's so random. I love it.黏土怪物跳舞？这也太随意了吧。我喜欢。
The prompt was random too: a group of cute 3D clay monsters joining a jumping retro dance party. And the agent built a moving demo in the browser.提示词本身就很随意：一群可爱的 3D 黏土小怪物加入一个跃动的复古舞池派对。然后智能体直接在浏览器里做出了会动的演示。
Okay, so it can see, understand tone, and build things. How do developers actually use it? Is it hard?好，所以它能看，能理解基调，还能搭建东西。开发者实际怎么用它？难吗？
Super easy. You just set model='deepseek-v4-flash-vision-exp' in the API. That's it.超级简单。你只需要在 API 里设置 model='deepseek-v4-flash-vision-exp' 就行。就这么简单。
Just change one line? And the price?就改一行？那价格呢？
Same as V4 Flash. Each image costs at most 384 tokens. So you pay by tokens, not by extra vision fee.和 V4 Flash 一样。每张图片最多算 384 tokens。所以你按 token 付费，没有额外的视觉费用。
384 tokens per image - that's pretty cheap. What about sending images?每张图片 384 tokens——挺便宜的。那怎么发送图片呢？
Three ways: you can paste base64 code, send a URL link, or use their new Files API.三种方式：你可以粘贴 base64 编码，发一个网址链接，或者用他们新的 Files API。
Files API? Is that new?Files API？是新出的吗？
Yes, just launched. You upload the image once, get a file_id, and use that file_id in many requests. No need to upload again. And it's free to upload.是的，刚刚上线。你上传一次图片，拿到一个 file_id，然后在很多请求里重复使用这个 file_id。不用重复上传。而且上传是免费的。
Oh, that saves time and bandwidth. Does the API work with different formats?哦，那省了时间和带宽。API 支持不同的格式吗？
Yes, three formats: Chat Completions, Messages, and Responses. So it fits into most agent tools easily.支持，三种格式：Chat Completions、Messages 和 Responses。所以它能很轻松地接入大多数智能体工具。
You know what I find interesting? This is an experimental model. Exp means they are still testing.你知道我觉得有意思的是什么吗？这还是个实验性模型。Exp 意味着他们还在测试。
Right, but it's already strong enough to make real PPTs and websites. Imagine what the final version will do.对，但它已经强到能做真正的 PPT 和网站了。想想最终版会做到什么程度。
And it mixes text and images together. That feels more human. We don't just read - we look.而且它把文字和图片混在一起处理。这感觉更像人类。我们不只是阅读——我们也在看。
Right. Before, AI was like a smart friend who could only hear your words. Now it can see your photos, your screenshots, your designs.对。以前，AI 就像一个只能听你说话的聪明朋友。现在它能看到你的照片、截图、设计稿。
So it's not about who talks the most. It's about who can see what you're pointing at.所以重要的不是谁话最多，而是谁能看懂你在指什么。
Exactly! And here's a funny thought - we spent years teaching AI to read our messy handwriting. Now we're teaching it to judge our PowerPoint taste.没错！还有个好笑的想法——我们花了好多年教 AI 认我们潦草的字迹。现在我们在教它评判我们的 PPT 审美。
Ha! My PPT taste definitely needs judging. So if listeners want to try it, just try that model name?哈！我的 PPT 审美确实需要被评判一下。所以如果听众想试试，只要试试那个模型名就行？
Yes, try model DeepSeek V4 Flash Vision Exp and see what it sees. And tell us - what would you ask an AI that can finally see?对，试试模型 DeepSeek V4 Flash Vision Exp，看看它能看见什么。也告诉我们——如果 AI 终于能看见了，你会让它帮你做什么？
We'd love to hear your ideas. See you next time on Learn English with Podcasts!我们很期待听到你的想法。下期 Learn English with Podcasts 再见！
See you next time!下期见！