跪拜 Guibai
← All articles
DeepSeek · AIGC · GitHub

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

By 程序员晓凡 ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Text-only models remain cheaper and faster than multimodal alternatives for many coding tasks. A two-cent-per-image relay sidesteps the need to switch models or wait for official vision support, keeping the cost advantage while removing a daily workflow bottleneck.

Summary

DeepSeek's text-only models cannot process images, a friction point for developers who want to paste error screenshots or design mockups directly into a chat. The open-source `claude-vision-skill` bridges this gap by piping images to a separate vision model—defaulting to Alibaba Qwen's VL models—and returning a text description that DeepSeek can read. Installation is a single prompt handed to Codex, which pulls the repo and writes the config automatically. After adding a Qwen API key, the skill works immediately without a restart. New users get 1 million free tokens from Alibaba Cloud Bailian, enough for roughly 7,000 image recognitions at about two cents each. In testing, the setup correctly read document screenshots, parsed a project startup error stack, and described front-end layouts with enough detail to generate matching code. The skill supports any OpenAI-compatible vision model, so users are not locked into Qwen.

Takeaways
DeepSeek's V4-Flash and V4-Pro models are text-only and reject image inputs outright.
The `claude-vision-skill` GitHub repo (1.9k stars) adds image understanding by routing images through an external vision model and returning a text description.
Default vision models are Alibaba Qwen's `qwen3.5-omni-plus` and `qwen-vl-max`, but any OpenAI-compatible vision API works.
New Qwen users get 1 million free tokens, covering about 7,000 image recognitions at roughly ¥0.02 per call.
Installation is a single prompt to Codex: it clones the repo, writes config, and sets up the skill automatically.
After adding a Qwen API key, the skill works immediately with no restart required.
Tested scenarios include reading document screenshots, parsing error stack traces, describing UI layouts for code generation, and translating English text inside images.
Qwen's vision accuracy is reliable for Chinese-language screenshots, documents, and interfaces, with occasional failures only on unusually complex images.
Conclusions

Piping images through a cheap external vision model is a pragmatic workaround that avoids the higher cost and latency of full multimodal models for the majority of text-heavy coding workflows.

The skill's design is model-agnostic at the vision layer, which means users can swap in a local vision model later and eliminate the API cost entirely.

Automating the install via Codex turns what would be a multi-step setup into a single natural-language command, lowering the barrier for developers who would otherwise skip the fix.

At two cents per image, the economics flip the usual tradeoff: it becomes cheaper to screenshot an error than to manually copy-paste and reformat a stack trace.

Concepts & terms
Multimodal model
An AI model that can process multiple types of input—typically text and images—within a single architecture. DeepSeek's V4 series is text-only, meaning it lacks native image understanding.
OpenAI-compatible API
An API that follows the same request/response format as OpenAI's chat completions endpoint. Many model providers adopt this standard so tools can swap models without changing integration code.
Agent / Codex
In this context, an AI coding assistant (such as OpenAI's Codex CLI or similar agent tools) that can execute shell commands, clone repositories, and edit configuration files based on natural-language instructions.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗