Add Vision to Codex: Connect DeepSeek-V4-Flash to qwen3-vl-flash (August 2026)

[//]: # (# Add Vision to Codex: DeepSeek-V4-Flash + qwen3-vl-flash Illustrated Tutorial (2026))

Want to add vision capabilities to Codex? When we swap Codex's base model to DeepSeek-V4-Flash — fast with powerful Agent capabilities — we run into a pain point: DeepSeek-V4-Flash is a pure text model and cannot see images by default.

In everyday programming, reading images is a real need — whether it's recreating a UI screenshot, analyzing a complex error popup, or double-checking a page layout. Relying on text descriptions alone is both inefficient and error-prone.

🎯 Core summary: This article will show you how to add vision capabilities to Codex in 3 simple steps in about 5 minutes. We'll combine the open-source plugin vision-skill with the cheap, effective Tongyi Qianwen qwen3-vl-flash model, giving your Codex the ability to "see images and write code" in an instant!

---

📚 Table of Contents (click to jump)

• Why Add Vision to Codex?
• How It Works: vision-skill + qwen3-vl-flash
• Before You Start
• Step 1: Get an API Key from Alibaba Cloud Bailian
• Step 2: Let Codex Install and Configure Automatically
• Step 3: Send an Image to Verify the Vision Feature
• FAQ

---

Why Add Vision to Codex?

After connecting DeepSeek-V4-Flash, Codex becomes a top-tier "code brain" — it can read and write, but it's "blind". If you only write pure backend logic, that's no problem; but as soon as you're doing frontend development and working from design mockups, it gets stuck.

Think of adding vision to Codex as a fine division of labor on a construction site:

• DeepSeek-V4-Flash is like a smart foreman, handling logical reasoning and writing code;
• the external vision model is a dedicated "quality inspector", responsible for understanding the image and translating it into a detailed text report for the foreman.

How It Works: vision-skill + qwen3-vl-flash

Thanks to the open-source community, there's already a mature, ready-made solution. It has two core parts:

1. vision-skill plugin: an open-source skill package (GitHub repo: asuojun/claude-vision-skill (opens new window)). It gives a base model without native image recognition a pair of "eyes" — it can read local or online images, call an external vision API, and feed the description of what it sees back to Codex.
2. Tongyi Qianwen qwen3-vl-flash: the "eye" itself we chose. It's designed for image-text multimodal (VL) scenarios and is more than enough for reading everyday UIs and screenshots. Most importantly: it's extremely cheap, and new users get a free quota. It's billed per token (input is only 0.15 RMB / million tokens), far more cost-effective than models charged per image.

Before You Start

Before you begin, make sure you have the following:

• [x] Basic setup done: we recommend first reading Connect Codex to DeepSeek-V4-Flash.
• [x] Alibaba Cloud Bailian account: if you don't have one, register for free on the Alibaba Cloud website.
• [x] Create an API Key: we'll do this in the next step.

---

Step 1: Get an API Key from Alibaba Cloud Bailian

First, we need to get the "key" for calling the vision model.

1. Log in to the Alibaba Cloud Bailian console (opens new window).
2. In the left navigation bar, find and enter API-KEY management.
3. Click to create a new key, and copy and save it.

⚠️ Security warning: Your API Key is like the password to your digital wallet! Never post it in public tutorials, group chats, or commit it to a public GitHub repository, or you risk having your quota stolen.

Step 2: Let Codex Install and Configure Automatically

This is the best step — you don't need to install anything yourself by typing code; just hand the job to Codex itself!

Open Codex, replace the API Key in the prompt below with yours, and paste it directly to Codex:

Please install asuojun/claude-vision-skill globally (https://github.com/asuojun/claude-vision-skill),
and configure it according to the instructions in the README.
For the vision model, please use qwen3-vl-flash from Alibaba Cloud Bailian.
The API Key is sk-xxxxxxxxx (replace this with the real API Key you just created).

Once Codex receives the instruction, it will automatically download the repository, check the local environment, and write the relevant configuration. The whole process is automated — just sit back and wait.

Step 3: Send an Image to Verify the Vision Feature

After the setup, how do you confirm that vision has successfully been added to Codex?

Simple: send a screenshot directly in the chat (for example, a web UI or an error screenshot) and ask it: "What's in this image?"

If it can accurately describe the content, colors, button positions, or the error code in the image, congratulations — your Codex has grown "eyes"!

(Note: if verification fails, go back and check whether the API Key in Step 1 was copied completely, or have Codex re-investigate the vision-skill configuration.)

---

FAQ

Q: Can't DeepSeek-V4-Flash really see images by itself? A: Correct — it's a pure text model. That's exactly why we need the external solution to add vision to Codex.

Q: Will this be expensive to use long-term? A: Very cheap. Vision models are billed per token. For qwen3-vl-flash, input is only 0.15 RMB / million tokens. Looking at screenshots during daily coding costs almost nothing, and new users even get a free quota.

Q: Do I have to use qwen3-vl-flash? A: Not necessarily. vision-skill uses the OpenAI-compatible standard format, so in theory any model that provides a vision API can plug in seamlessly. This article recommends qwen3-vl-flash mainly because of its excellent value — perfect for reading UI mockups in everyday work.