DeepSeek-OCR is INSANE (2.3M Downloads, $0)

Show Notes

Watch the video: https://www.youtube.com/watch?v=0US4LMIB0Vk

DeepSeek-OCR shipped free open weights on Hugging Face and quietly repriced the entire document-API line on your P&L.
Playbooks + my answers, free - Church of SEO Jesus: https://www.skool.com/church-of-seo-jesus

What you get in the next 7 minutes:
- The $150 vs $17 monthly math on 100k document pages (Textract vs a rented L4)
- Why DeepSeek-OCR at 3.3B params fits on a $0.60/hour GPU
- The one-afternoon brief to send your dev or agency Monday morning
- The workflows to KEEP on Textract (handwriting, dense tables, regulated data)
- Why two releases in the collection means this is a maintained line, not a one-off

Open it yourself: https://huggingface.co/deepseek-ai
AWS Textract pricing: https://aws.amazon.com/textract/pricing/
Google Document AI pricing: https://cloud.google.com/document-ai/pricing
Azure Document Intelligence pricing: https://azure.microsoft.com/en-us/pricing/details/ai-document-intelligence/

00:00 The 2.3 million download number your vendor won't email you about
00:14 What DeepSeek-OCR actually is (3.3B params, 6GB, no gate)
00:47 Textract vs DeepSeek-OCR: the $150 to $17 monthly math
01:43 Why OCR is the small door (vision workflows next)
02:03 Subscribe + Church of SEO Jesus
02:34 GPT-4o vision: the bigger line DeepSeek also covered
02:58 Live proof: hitting the Hugging Face API and the P&L calc
04:02 Open the download counter yourself
04:14 The honest part: procurement, compliance, and what to keep on Textract
05:19 The Monday-morning brief for your dev
06:00 DeepSeek-OCR-2 is already out (this is a maintained line)
06:42 Your competitor's technical cofounder already deployed it
07:20 Tomorrow: the transcription bill a free model just erased

Audio edition of the Stewart Vickers AI YouTube video of the same name. Full transcript on this episode page.

What is Stewart Vickers AI?

The audio edition of the Stewart Vickers AI YouTube channel. Each episode takes one AI release - a model, a tool, a price cut - and translates it into money for business owners: what got cheaper, who got undercut, and what to do about it this week. Operator credibility over theory, specific numbers over vibes. Every episode links the original video and the free Church of SEO Jesus community where the prompts and workflows live.

Look at that number. 3 million people already downloaded a free Modal that eats what your document API bills you for, and your vendor is not going to send you an email about it. 3 billion parameters, 6 gigabytes on disk, small enough to fit on a rented GPU that costs less than a coffee an hour. If you push invoices, receipts, contracts, IDs, or scanned PDFs through Textract, Google Document AI, or Azure Read, you are paying a meter. The meter just became optional. Here is the math that decides whether this is a real story or a headline.

100,000 document pages a month is a normal small business volume. A busy accountancy, a mid-sized logistics operator, a property management firm scanning tenant paperwork. 50 PR thousand pages, $150 a month, $1,800 a year. Boring but growing. Now DeepSeek. 60 an hour.

16. Same job, $132 saved every single month on one workflow in one small business. Multiply across every operator running this quietly. An OCR is the small door. DeepSeek shipped this under a Vision Modal umbrella, meaning the same weights that pull text off a scan can read a chart, describe a photo, classify a product image. Every workflow you were pointing at GPG-4o Vision just got a free, open alternative to benchmark against.

I am Stuart. Every Modal release like this one lands on this channel the day it ships, so subscribe and the next one finds you instead of you finding it three months late on an invoice. The pinned comment has the exact commands and the Modal link. And the Church of SEO Jesus, my free school community, is where the full swap-out playbook lives and where I answer when your devs first deploy breaks. Link below. Here is the second door.

GPG-4o Vision on a full page bills somewhere between half a cent and a cent and a half PR page, depending on detail. A hundred thousand pages a month lands you between $500 and $1,500 before retries. Same L4 handles it for $17. The Vision line was quietly larger than the OCR line and DeepSeek covered both. Watch this. 38 million downloads, public, free, not a preview, not a waitlist, not a locked license page.

21 gigabytes, which fits comfortably on a 24 gig L4 or a consumer R-TX4090. Cheap infrastructure, your dev already knows how to spin up. Then I asked a calculator Modal to sanity check the PNL at a hundred thousand pages a month. Hosted bill $150. 16. 84.

Five numbers, whole story. And I gave the calculator the current public rates, not an optimistic scenario, run pod L4 pricing, real throughput. If anything, the math is generous to the incumbents because I did not price in T-extracts PR feature upcharges for tables, forms, and handwriting, which is where the meter actually runs hot. co slash deep CK yourself right now. 3 million. Proof, not me telling you.

Now the honest part, because I want you to actually use this and not blow your foot off. Deep seek OCR is a Chinese lab release. If you sell into regulated buyers who write your vendor list, the word Chinese alone kills the conversation. Fine. Run it on workflows. They never audit internal invoice matching vendor, onboarding marketing receipts.

The stuff hosted AP is quietly print money on handwriting, dense tables, low contrast scans, and multilingual pages are historically where open OCR loses to T-extract. Keep those on the paid API, migrate the 90% of your volume that is clean printed PDFs, and watch a five-figure annual bill shrink to a four-figure one. Compliance. If you are hip, a PCI, or a European operator with a signed data processing agreement, do not tear it up for a savings line. Self-hosted has its own compliance story to write. Run parallel first, prove the outputs match on your own document set, then move only when your lawyer nods.

Here is the sentence you send your developer or your agency Monday morning. Deploy the deep seek OCR Modal from hugging face on a rented L4 or A10G. Wrap it in a small API that takes a PDF and returns text and bounding boxes. Point one non-regulated workflow at it. Log the accuracy against our current T-extract output for two weeks. Report back the diff.

One afternoon of dev time. A two-week comparison. An answer that either saves you $1,500 a year on this one workflow or tells you the paid meter earned its keep. Both outcomes are cheap to buy. If your agency quotes it as a two-week project, get a second opinion. Now here is the part you should be nervous about.

Already two deep seek OCR releases in the collection. It is not a one-off, it is a maintained line. The version after this one gets better, the version after that runs faster. The compounding curve is not on the vendors selling you PR page pricing anymore, it sits on the freeweights on hugging face. I got this story from watching the download counter tick. 3 million in the window between writing the script and recording it.

Every one of those downloads is somebody spinning up an instance. Most of them are hobbyists. Some of them are your competitor's technical co-founder. Your competitor in your vertical who has one technical friend just cut their extraction costs to a GPU hour. They are not going to publish a case study about it. You find out when they underprice you on a bid you used to win comfortably.

Which is why the Church of SEO Jesus exists. It is my free school community, link in the description. The spin-up brief I gave my own dev, the L4 rental walkthrough, the two-week t-extract accuracy log template, and the vendor swap-out playbooks all sit inside. Where I answer when your first deploy hits a wall, free to join. The window on this one is short, not the Modal, the Modal will still be free next year. The window is the arbitrage.

When your competitor moves first, PR-page OCR pricing stops being a cost line and starts being a positioning problem. The operator running local weights has structurally cheaper unit economics than you do on that workflow forever. Tomorrow I am running the same math on a different vendor line that a free open Modal just erased, and it involves your monthly transcription bill. Point at the end screen.