Give DeepSeek Eyes with an Open-Source Vision Skill That Costs Pennies

1. DeepSeek is cheap and easy to use, but it's "blind"

More and more people are using DeepSeek. The reason is simple: it's cheap and keeps getting stronger.

As soon as the official V4-Flash version came out, performance jumped again, and just a few days ago, the official V4-Pro version was also released.

Some platforms have disclosed that DeepSeek token consumption has reached 8 trillion, showing that people are genuinely using it with real money, not just talking about it.

However, those who have used it for a while have likely hit the same stumbling block: it can't see images.

If you send it a screenshot, it's clueless and directly tells you, "I cannot process images." This doesn't matter much for chatting, but when it comes to real work, it's a bit frustrating.

Take the most common example. Your code throws an error, and you want it to take a look. You send a screenshot, but it can't read it. You have to copy and paste the error message word by word into text, and you also need to clearly describe which file and which line. When you encounter a messy error stack trace, the copy-pasting can make your head spin.

Another example is front-end development. The design mockup is right in front of you, and you want it to write the interface accordingly. It can't see the image, so you can only describe it verbally: a button on the left, an input box on the right, the color is blue... It takes a lot of effort, and the description still feels different from what you have in mind.

In these scenarios, the bottom line is that it has no "eyes"; it is not a multimodal model.

Actually, it's not just DeepSeek; many text-focused large models have this shortcoming. It's just that DeepSeek has many users and is cheap, so people are especially willing to use it for work, which amplifies this flaw.

Fortunately, there is now a ready-made way to fill this gap, and the cost is negligible. Let's get into the details below.

2. An open-source Skill solves the problem

You don't need to wait for an official update for this capability. The open-source community already has a solution ready, named claude-vision-skill.

Don't be put off by the name "Claude." Although it was made for Claude, it can be installed on any Agent, and the Agent will adapt itself after installation.

This Skill currently has 1.9k stars.

GitHub address: https://github.com/asuojun/claude-vision-skill

The principle behind it is quite simple: first send the image to a model that can recognize images, let that model describe the content of the image in text, and then feed this text back to DeepSeek. DeepSeek receives text, so it naturally "understands" it.

To put it plainly, it's like hiring a "sighted" model as a translator to turn images into text and then relay it to DeepSeek. It doesn't need to be able to see images itself; it just needs to be able to read the translation result.

So you need to additionally configure a model that can recognize images.

The repository recommends Alibaba Qwen's qwen3.5-omni-plus and qwen-vl-max by default.

Why Qwen? Because it's cheap. Alibaba Cloud Bailian gives new users 1 million tokens for free, which works out to about two cents per image recognition, practically free.

Of course, if you have other vision models that support OpenAI-compatible APIs, you can use them too; it's not picky. This Skill essentially does a "porter" job, and the vision model in the middle can be swapped for any other.

3. Installation

Just throw the following content to Codex, and it will handle the rest on its own; you don't need to manually type commands or modify configurations.

Follow the README at https://github.com/asuojun/claude-vision-skill to help me configure image recognition.

The whole process will pull the code and write the configuration by itself; you just watch.

After installation, there's one last step: configure the API Key for the vision model.

By default, it uses Alibaba Qwen. You need to go to the Qwen AI platform: https://platform.qianwenai.com/home/api-keys to create an API Key.

The 1 million free tokens for new users can recognize images about 7,000 times.

Even if you use it every day, it's enough to last for several months. When the free quota runs out, continuing to top up isn't expensive either; it's two cents per use, which you basically won't feel in daily use.

Two small things are worth noting. First, the first time you use it to recognize an image, the Agent might pop up a prompt asking you to confirm calling an external model; just click agree.

Second, don't casually share your API Key. Qwen charges by usage, and although it's cheap, it would be a loss if someone with bad intentions used it to rack up charges.

4. How effective is it after installation?

It can be used immediately after installation, no restart needed, no need to fiddle with it again.

I casually tried it with an image and asked it to read the content of the image.

The result was quite good; the text and structure were basically all recognized.

I also threw a screenshot of a ruoyi project startup error at it.

With this, DeepSeek has "opened its eyes." From now on, when you encounter an error while coding, just send a screenshot directly; for what an interface looks like, one image explains it all.

I basically use it as the default configuration now. When modifying the front end, I throw a screenshot over, and it can describe the layout, colors, and spacing with high accuracy, saving a lot of back-and-forth when coding accordingly.

Reading document screenshots, reading errors, and even translating English in images can all be done conveniently. Previously, I needed to copy the text into it; now, one image does the job.

As for the accuracy of the vision model itself, Qwen's set performs quite stably in Chinese scenarios, and is basically sufficient for daily screenshots, documents, and interfaces. If it encounters a particularly complex image, it might occasionally fail, but that's an isolated case and doesn't affect the overall picture.

Summary

Adding image recognition capability to DeepSeek costs almost nothing, but the effect is immediate. The principle is not complicated: just borrow a model that can see, turn the image into text, and forward it.

Installation is just a matter of one command, taking a few minutes from start to finish. For those who need it, it's worth a try.