How to Give DeepSeek Eyes: Adding Vision to a Text-Only Model in 10 Minutes
**AI Export Lab** | July 2026...
**AI Export Lab** | July 2026
DeepSeek V4 is one of my favorite tools. Fast, cheap, strong at coding. It has one loud weakness for anyone doing catalog or document work:
**It can't see images.**
For product catalogs this is a real problem. About half of what I care about lives inside pictures — OEM numbers stamped on drawings, dimension tables rendered as flat images, part codes printed on photos of the actual part.
The fix turned out to be smaller than I expected. It's not a framework. It's not a plugin. It's basically one instruction and one API key.
---
The Idea In One Sentence
You don't need one AI that does everything. You need one AI that knows how to call another AI when it needs a skill it doesn't have.
DeepSeek is very good at reasoning about text. Vision models are very good at reading images. Wire them together and DeepSeek behaves as if it can see.
---
The Two APIs I Actually Use
In mainland China:
- **Doubao Vision Pro 1.5** — my workhorse. Fast, cheap, reads Chinese and English text inside images accurately. This handles my volume work.
- **Qwen-VL** — my backup. I use it when I want a second opinion on a hard image, or when Doubao gives an answer that doesn't smell right.
Both give a free monthly quota. For a few hundred images per month I pay nothing.
If you're outside China, the same shape exists in several places:
- **Qwen-VL** through Alibaba Cloud International (English dashboard, foreign card, free tier)
- **Xiaomi's overseas AI platform** (Google account sign-up, no Chinese phone needed)
- **Kimi** and **GLM** through OpenRouter — one account, no Chinese phone
- **Gemini, Claude, or GPT-4o** — vision is built-in, more expensive per image but zero setup friction
Pick the one you can sign up for. They all fit the same wiring.
---
The Wiring Is Almost Nothing
Here's what you actually do:
- Sign up for the vision provider. Copy your **API key** and your **endpoint URL**.
- Tell your text AI: *"Whenever you need to read an image, call this API. Here's the key. Here's the URL."*
- That's it.
No orchestration platform. No plugin marketplace. One file that says "here's a tool called `read_image`, this is how you call it."
Inside your automation script, the flow looks like:
- DeepSeek receives a job that involves an image.
- DeepSeek returns: *"I need to read this image."*
- Your script intercepts that request, calls the vision API, gets back OCR or a description.
- Your script feeds that text back to DeepSeek.
- DeepSeek continues reasoning with the image content now available as text.
You just built a two-model system without any framework.
---
Make It Reusable: Turn It Into a Skill
If you're only doing this once, hard-code the API key and move on.
If you know you'll use this pattern in multiple projects — and if you're the kind of person reading this post, you will — turn the wiring into a **skill**.
A skill is just a small folder or file with:
- A short description of what capability it provides ("read images")
- The API details (key, endpoint, model name)
- A usage note in plain English
Now when you start a new project, you tell your AI "load the vision skill" and the wiring is instant.
I have skills for vision APIs, TTS voices, search APIs, and image generation. Each one turns a repeated setup task into a single sentence.
This is the real productivity gain from AI tooling. Not writing code faster. Not writing zero code. **Moving setup from "do it every project" to "do it once, ever."**
---
What It Actually Costs
Short version, with real numbers from a batch job I ran a few weeks ago:
- A few hundred images per month: **free**, under the vision provider's free tier
- Around 6,000 images in one batch: **about $5** (I wrote the full breakdown [here](/ai-export-lab/6000-images-for-5-dollars-doubao-vision-pro/))
- Tens of thousands per month: still cheaper than one hour of a data entry team's time
The image bill is not what breaks a project. What breaks a project is complexity. Which is why keeping the wiring simple — one API, one skill file, one clear rule for when to call it — matters more than picking the "best" vision model.
---
When to Use This vs a Native Multimodal Model
Native multimodal models — Claude, Gemini, GPT-4o — can see images directly. Why bother with the wiring at all?
You often shouldn't. Depends on volume.
- **Small, high-value job.** One-off analysis of a few dozen images. Use a native multimodal model. Skip the split.
- **Large batch job.** Thousands of similar images with a consistent format. Split the work: cheap vision model for the read, cheap text model for the reasoning. Total cost drops a lot.
The economics flip at scale. Native multimodal models charge you a premium because they can do everything. When you know exactly what you need — "read the text on this image and return JSON" — mixing a cheap vision model with a cheap text model is dramatically cheaper.
Same principle as the [manager-worker pattern](/ai-export-lab/claude-code-deepseek-manager-worker-pattern/). Match the task to the smallest tool that handles it well.
---
Where This Wiring Doesn't Help
To be honest about the failure cases:
- **The image is truly ambiguous.** A blurry photo where even a human isn't sure what's on it. Vision AI will confidently return something wrong. You'll spot-check and go "no, that's a Volvo, not a Scania." Human review still needed.
- **The image requires context outside the image.** "Is this the part number for the 2015 model or the 2016 model?" Vision AI only sees the picture; it doesn't know your business.
- **The image is in a language your vision model wasn't trained on.** Doubao and Qwen are strong on Chinese and English. For Arabic, Thai, or handwritten Russian, test on a small sample before committing to a batch.
Adding vision to a text AI unlocks a huge amount of work. It doesn't turn AI into a magic OCR-for-anything.
---
What This Unlocks
Once your text AI can see, whole categories of work open up:
- Old paper catalogs that only exist as scans
- Product photos from a wholesale supplier's website
- Screenshots from a competitor's spec sheet
- Handwritten notes from a supplier visit
- Damage inspection photos from a customer complaint
Every one of these is data your business already has and probably isn't using because it's trapped in image form.
Adding a vision API is the difference between "we have this data somewhere" and "this data is now searchable and structured."
For a solo builder trying to compete with bigger companies, that shift is significant.
---
**Have you wired up a vision API this way?** Curious what people are using it for outside of catalog work. Reply on the [contact page](/contact/) if you've tried it for something unusual.
*This post is part of the cornerstone series [How I Turned a 600-Page PDF Into a Product Database for Under $10](/ai-export-lab/how-i-turned-a-600-page-pdf-into-a-product-database-for-under-ten-dollars/).*