Research + corpus-building uses. Benchmarks don’t count.
The drawers
A tool can be filed in more than one drawer
DRAWER I
AI & History
Machine learning, vision-language models and speech models. Every card here carries notes on where your sources go, how it was tested, and how it fails.
Guide cards
Counts research and corpus-building uses in the window, plus “I’m using this now” check-ins. Benchmarks don’t count.
tool · No. 018AI
DeepSeek-OCR
Small open model that compresses page images for fast bulk conversion. Built mainly as research into compression rather than careful transcription.
At a glance
Maker
DeepSeek AI
Released
Oct 2025 (v2: Jan 2026)
Licence
Open weights (MIT; v2 Apache-2.0)
Cost
Free; third-party hosting
Runs
Your own GPU or third-party APIs
Skill
Programming
Historian’s notes
Where your sources goNo official hosted service; self-hosting keeps sources local.
How it’s been testedMedian error 6.7% on Ancient Greek editions and 20.8% on Arabic (Inria 2026).
Known problems on historical sourcesGuesses from language rather than reading the page: it “corrects” unusual words to ones that aren’t there and loops on long outputs. Among the weakest on degraded microfilm.
Used in research0
No published research uses found yet. Using it in a project? Tell us.
Nobody has said they’re using it in 2026 yet.
“I’m using this now” asks you to sign in with ORCID.
One count per person per year; it adds to Trending, and names are never shown.