Last Wednesday at 2am, I snapped a photo of my handwritten ledger. The AI pulled out the amounts in 3 seconds. Not the future—this week.

What is this thing?

DeepSeek quietly shipped a "vision" model called v4-flash-vision-exp. Simple version: you send it a photo, it reads it like a human—text, objects, answers.

Who's using it? My buddy Ajie runs Xiaohongshu (Little Red Book) account ops in Hangzhou. He processes 50+ product photos a day. Used to pay an assistant ¥3,800/month just to write product descriptions. Now the AI drafts them, the assistant just polishes. "The money I save covers three months of rent," he told me in a coffee shop near West Lake, dark circles under his eyes.

I made this mistake too: I assumed "vision AI" meant GPT, VPNs, foreign cards, off-shore servers. Nope. Our homegrown DeepSeek takes WeChat Pay.

Cost to replicate today

  • Money: ~¥0.1–0.3 per image (varies by size and complexity); 1,000 images/month runs about ¥100–300
  • Time: 30 minutes from signup to first successful API call
  • Tech barrier: You can copy-paste a long string that looks like a password (the API key). Knowing it's a "pass" is enough
  • First step: platform.deepseek.com → scan WeChat to log in → top up (¥1 minimum) → docs page → find the "Vision" section → copy the sample code

Note: this isn't for everyone. If you only process 5 images a day, manual is faster.

Is it for you?

  • Just starting (under 10 images/day): I'd tell you to stay manual. Wait till volume grows
  • 1–2 steady clients: Worth a try. Saves 5–8 hours a week
  • Scaling (50+ images/day): No-brainer. Wire it up tonight. I waited until clients were chasing me—regretted it

No pressure. Bookmark it. Come back when you're drowning.