UPGPTs

English

中文

English

中文

Appearance

Add Vision to Codex (2026)｜vision-skill + Alibaba Qwen qwen3-vl-flash, Step by Step ​

In Connect Codex to DeepSeek-V4-Flash, we gave Codex DeepSeek-V4-Flash as its underlying model. It's fast and a much stronger agent, but it has one weakness: DeepSeek-V4-Flash isn't multimodal — it can't see images.

Looking at images matters a lot in coding — UI screenshots, error messages, page layouts. Describing them in text is slow and error-prone.

TL;DR: Adding vision to Codex takes three steps — ① create an API Key in the Alibaba Cloud Bailian console; ② hand the open-source repo asuojun/claude-vision-skill and the API Key to Codex and let it install and configure itself; ③ send an image to verify. Vision model: Alibaba Qwen qwen3-vl-flash — cheap, and new users get a free quota. About 5 minutes.

Contents:

  1. Why Codex Can't See Images
  2. What vision-skill and qwen3-vl-flash Do
  3. What to Prepare Before You Start
  4. Step 1. Create an API Key in the Bailian Console
  5. Step 2. Hand the Repo and API Key to Codex
  6. Step 3. Verify It Sees Images
  7. FAQ
  8. Related Reading

Why Codex Can't See Images ​

After connecting DeepSeek-V4-Flash, Codex's underlying model is text-only: it can read and edit code, but it can't process images. That's fine for pure coding, but it falls apart the moment you need to work from screenshots — recreating a UI, reading an error dialog, checking a design against the mockup.

Think of it like a construction site: Flash is the foreman, reasoning and editing code, and you hire a dedicated person whose only job is to look at images and translate what they see into text. The main model thinks; the vision model looks.

How the reasoning model and vision model split the work in Codex

What vision-skill and qwen3-vl-flash Do ​

Luckily, someone already open-sourced the solution: vision-skill, which gives Codex a pair of eyes.

  • vision-skill: an open-source skill pack that attaches "eyes" to a main model with no native image support — it reads local or online images, calls a vision API, and puts the description back into the conversation. Repo: asuojun/claude-vision-skill.
  • Vision model: the part that actually looks at the image. I picked Alibaba Qwen's qwen3-vl-flash — a vision-language model built for reading images, good enough for everyday UI and screenshots, and cheap.

Alibaba Qwen qwen3-vl-flash model overview

Alibaba Qwen qwen3-vl-flash pricing

qwen3-vl-flash is billed per token, not per image (input ¥0.15 / 1M tokens, output ¥1.5 / 1M tokens), and new users get some free quota.

What to Prepare Before You Start ​

  • ☐ Finish the previous guide: Connect Codex to DeepSeek-V4-Flash. You can still follow this one without it, but the setup matches the demo environment
  • ☐ An Alibaba Cloud Bailian account: register if you don't have one
  • ☐ An API Key: create it in the Bailian console (next step shows how)

Step 1. Create an API Key in the Bailian Console ​

Sign in to the Alibaba Cloud Bailian console, open API-KEY management, create a key, and copy it. You'll need it in a moment.

Create an API Key in the Alibaba Cloud Bailian console

Security reminder: this key is your money. Use it only on your own machine — never paste it into a public tutorial or commit it to a public repo. Anyone who gets it can spend your quota.

Step 2. Hand the Repo and API Key to Codex ​

No need to download anything or hand-edit config. Just paste this into Codex and let it follow the README itself:

markdown
Install asuojun/claude-vision-skill globally (https://github.com/asuojun/claude-vision-skill),
and configure it according to the README.
Use Alibaba Cloud Bailian qwen3-vl-flash as the vision model.
The API Key is sk-YOUR_API_KEY (replace it with the one you just created).

Note: replace sk-YOUR_API_KEY with the key you created in the Bailian console — don't copy the example as-is.

Codex will download the repo, check the environment, and write the config for you — no manual commands.

Codex auto-installing and configuring vision-skill

Step 3. Verify It Sees Images ​

Once configured, paste a screenshot and ask Codex to describe it. If it correctly says what's in the image, the eyes are installed and vision works.

If verification fails, go back to Step 1 and double-check the API Key, then ask Codex to check the vision-skill config.

FAQ ​

Q: Can't DeepSeek-V4-Flash see images? A: Correct. It's a text-only model — not multimodal — so it has no native image recognition. You add a vision model on top to fill that gap.

Q: Does adding vision to Codex cost extra? A: The vision model is billed per token, not per image. For qwen3-vl-flash: input ¥0.15 / 1M tokens, output ¥1.5 / 1M tokens. New users also get some free quota, so everyday image reading is very cheap.

Q: What is vision-skill? A: An open-source skill pack that attaches "eyes" to a main model with no native image support — it reads local or online images, calls a vision API, and puts the description back into the conversation. Repo: github.com/asuojun/claude-vision-skill.

Q: Do I have to use qwen3-vl-flash? A: No. vision-skill uses an OpenAI-compatible format, so you can swap in any vision-capable model. This guide uses qwen3-vl-flash because it's cheap, gives new users free quota, and is plenty for everyday UI and screenshots.

Q: How do I confirm vision is working? A: After configuring, paste an image and ask Codex to describe it. If it describes the image correctly, vision is live.