Give DeepSeek eyes — route images through Alibaba Cloud Bailian vision models so models that can't natively see images can still understand them.
DeepSeek (and many other LLMs) cannot natively see images. When you share a screenshot, photo, or diagram, the model literally can't process it.
DeepSeek has hinted at a multimodal model coming soon — but until then, this tool fills the gap.
Your Image → eyes.py → Bailian Vision Model (Qwen series) → Chinese Text Description → Your Chat Model
The script sends your image to Alibaba Cloud Bailian's vision API, gets back a detailed description, and prints it to stdout. The LLM reads this text and answers your question — it never needs to "see" the image directly.
| Model | Use case | Precision | Speed |
|---|---|---|---|
qwen3-vl-plus (default) |
Always use unless you need speed. Photos, diagrams, small text, detailed scenes. | ★★★ | ★★ |
qwen3.6-plus |
Legacy flagship. Use when vl-plus is unavailable. | ★★ | ★★ |
qwen3.6-flash |
When you just need a quick look. Simple photos, casual use. | ★ | ★★★ |
Real test (complex illustration, all with --high-res):
| qwen3-vl-plus | qwen3.6-plus | qwen3.6-flash | |
|---|---|---|---|
| Output detail | ~1200 words | ~500 words | ~400 words |
| Hidden text found | "LOVE" on balloon | none | none |
| Artwork identified | WLOP "The Sky Garden" 2018 | no | no |
| Color errors | none | none | rainbow → "yellow" |
Recommendation: Default
qwen3-vl-plus --high-rescovers 90% of use cases. Switch to--model qwen3.6-flash --high-reswhen you just need a quick look.
Alibaba Cloud Bailian gives new users free API quota — no payment needed to start:
- 1 million tokens per model series
- Valid for 90 days from activation
- Mainland China region only
- Tip: enable "Stop when free quota exhausted" in the console to avoid unexpected charges
After the free quota runs out, vision models start at ~¥1 per million tokens.
What you need:
pythoninstalled. Runpython --versionin terminal to check.
The skill is two files: SKILL.md + eyes.py. Download them into Claude Code's skills folder.
macOS / Linux:
mkdir -p ~/.claude/skills/deepseek-eyes
curl -o ~/.claude/skills/deepseek-eyes/SKILL.md https://raw.githubusercontent.com/hawkongz/deepseek-eyes/master/SKILL.md
curl -o ~/.claude/skills/deepseek-eyes/eyes.py https://raw.githubusercontent.com/hawkongz/deepseek-eyes/master/eyes.pyWindows (PowerShell):
New-Item -ItemType Directory -Force -Path "$env:USERPROFILE\.claude\skills\deepseek-eyes"
Invoke-WebRequest -Uri "https://raw.githubusercontent.com/hawkongz/deepseek-eyes/master/SKILL.md" -OutFile "$env:USERPROFILE\.claude\skills\deepseek-eyes\SKILL.md"
Invoke-WebRequest -Uri "https://raw.githubusercontent.com/hawkongz/deepseek-eyes/master/eyes.py" -OutFile "$env:USERPROFILE\.claude\skills\deepseek-eyes\eyes.py"To update later: re-run the same commands.
pip install dashscopeRegister at Bailian Console, create an API key, then save it to your system:
macOS / Linux — run in terminal:
echo 'export DASHSCOPE_API_KEY="your-api-key"' >> ~/.bashrc
source ~/.bashrcUse ~/.zshrc if you're on zsh.
Windows — run in PowerShell:
[Environment]::SetEnvironmentVariable("DASHSCOPE_API_KEY", "your-api-key", "User")Restart terminal after setting. For temporary use: $env:DASHSCOPE_API_KEY="your-api-key" (current session only).
Share any image in Claude Code and ask a question about it — the skill handles the rest.
Drop SKILL.md and eyes.py into ~/.claude/skills/deepseek-eyes/. See Quick Start for the one-liner. Make sure pip install dashscope and DASHSCOPE_API_KEY are set first.
Run eyes.py directly — it's a standalone script. Feed the output into any LLM conversation.
├── README.md # You are here (English)
├── SKILL.md # Skill definition
├── eyes.py # Core script — image → vision model → text
├── requirements.txt # Python dependency (dashscope)
├── LICENSE # MIT
└── zh-CN/
└── README.md # 简体中文
deepseek vision multimodal claude-code skill bailian qwen image-understanding
Contributions welcome. See CONTRIBUTING.md for guidelines.
MIT