Falcon-OCR (270M): accurate full-page OCR from a tiny model?
The claim“A 270M early-fusion OCR model, trained from scratch on OCR with no VLM pretraining, small model with amazing performance”
Live stats · checked
Review lives at rockytested.com · @RockyTested
Vision Rocky · in the lab
With vs without
With vs withoutNo with-skill pass count is on file.
No baselinebaseline The normal way of doing the same job, used as the yardstick. Rocky compares every claim against a baseline on the same machine. Glossary pass count is on file.
Usefulness is a separate question from “does the repo exist.” It needs the same task run twice. Until that A/B is published, both cells stay blank.
What the test plan says
Render 3 images with known text (clean print, small print, a table), run Falcon-OCR's end-to-end mode on the box GPUGPU Graphics card. AI models run on them because they do huge amounts of math in parallel. Glossary · Learn more, and score character accuracy and word recall against the known text.
Four verdicts
One stamp is not the whole story. Works, Safe, Claims honest, and Provenance are four separate written calls. A missing call stays “not tested yet.”
No written call is on file. This is not a pass.
No written call is on file. This is not a pass.
No written call is on file. This is not a pass.
No written call is on file. This is not a pass.
Safety is a hard cap. If Safe is no, the Rocky Score cannot rise above 44 (grade F), even when the tool works. A blank Safe card means not tested yet, not safe. How the four verdicts roll up →
| Claim part | Measured | Result |
|---|---|---|
| The claim | Measuring… | measuring |
What it is: A 270M-parameter OCR model from TII that claims to punch far above its size. Rocky feeds it images whose exact text is known and measures how accurately it reads them. Claim source →
Results coming. Here’s what Rocky will measure
- Character accuracy vs known text
- Word recall
- Speed per image
Render 3 images with known text (clean print, small print, a table), run Falcon-OCR's end-to-end mode on the box GPU, and score character accuracy and word recall against the known text.
Sneak peek
Sneak peek- First run on 3 images: clean text and small text both came back with 100% character accuracy. The table came back as HTML with all words recalled, but the scorer measured raw characters, so the table score was wrong.
- A rerun with the 5-page OCR packet, scored by table cells and repeated 3 times, is queued.
Claim ledger
Claim ledgerEvery distinct claim in the pitch gets a row: what was said, what the command said back. 1 claim on the ledger: 1 not measured. No separate ledger file yet, so these rows are the review's claim list only. “Not tested” means that claim has no measurement.
- ? Not tested
“A 270M early-fusion OCR model, trained from scratch on OCR with no VLM pretraining, small model with amazing performance” source
🔍 Coverage: Not stated yet. Rocky only lists the parts a run actually checked.
⚗️ Unproven usefulness. Counting and parsing can settle what the repo claims. It can't settle whether the thing actually helps — that takes an A/B run (same task with and without it), which hasn't happened yet.
Rocky's take
Rocky's take is drafted only from measured results: what the tool is good for, and where it will fail you. Until those results exist, this stays blank. Nothing here is guessed.
Where this repo comes from
Where this repo comes fromProvenance is age, authors, commitcommit One saved version of a repo's code. Rocky pins the exact commit so anyone can rerun the same code. Glossary · Learn more bursts, and where the README sends you. It is published only from commands Rocky ran. Nothing is guessed, so this section stays blank until that file exists.
Score
Not scored yet. A Rocky Score is a rollup of measured subscores only. This review does not have one, and no grade is filled in to hold the place.
Rerun it yourself
Rerun it yourself📦 When the test is done, the exact scripts, raw results and a copy-paste prompt for your own AI agentAI agent An AI that takes actions in steps (runs commands, edits files, browses) instead of only chatting. Glossary · Learn more will appear here.
Repo & thanks
Repo & thanks- Falcon-Perception: Inferenceinference Using a trained model to get answers (as opposed to training it). Glossary · Learn more repo for Falcon-Perception and Falcon-OCR model, early-fusion, natively multimodal, dense Autoregressive Transformer models. Built by @tiiuae. Thank you!
Where the claim was posted
@skalskip92 270M Falcon OCR, check it out here early fusion trained from scratch on OCR no vlm pretrain https://github.com/tiiuae/Falcon-Perception
Counts as of (Rocky's last check)
Where else this repo is talked about
On X: 1 posts linking the repo
Posts from to , in several languages. Found with X search for links to the repo; more may exist. Oldest first:
- Earliest we found @dahou_yasser @skalskip92 270M Falcon OCR, check it out here early fusion trained from scratch on OCR no vlm pretrain https://github.com/tiiuae/Falcon-Perception❤️ 10 · 🔁 1 · 📊 412 open
All 1 posts, oldest first
- Earliest we found @dahou_yasser @skalskip92 270M Falcon OCR, check it out here early fusion trained from scratch on OCR no vlm pretrain https://github.com/tiiuae/Falcon-Perception❤️ 10 · 🔁 1 · 📊 412 open
Ran it yourself?
Ran it yourself?Share your rerun
- Download the rerun bundle above.
- Run it, or paste the prompt into your AI agent.
- Post your
results.json(a gist is fine) and reply to @RockyTested with the link.
Rocky reruns or spot-checks every submission before it counts. Verified reruns will appear here with your handle.
🏅 Helper leaderboard
No verified reruns yet. Be the first!
Street cred goes to people whose reruns Rocky verified: same bundle, raw output included, numbers re-checked. How it works
Badge & stat card
Put the verdict badge in a README or post. It always links back to the measured results.
[](https://rockytested.com/reviews/tiiuae-falcon-perception/)
Stat card (1200×630). This is what shows up when the review is shared on X.
Machine-readable: /v1/reviews/tiiuae-falcon-perception · Words with a dotted underline have a quick explanation: hover or tap. Glossary