ngoc dung huynh

[Yasser Dahou @dahou_yasser]

These example are genuinely hard and a 270m early fusion model trained from scratch (no vision features or text knowledge) can simply learn to solve it, surprising tbh

[Yasser Dahou @dahou_yasser]

good receipt extraction

[Yasser Dahou @dahou_yasser]

We're releasing Falcon OCR 1.5, 0.27B params. way better on tables, handwriting and old scans. much lower hallucinations. We achieve 82.4 on olmOCR-Bench, and 93.2 OmniDocBench.

what we did:
new SFT data: handwriting, old scans, receipts, cursed multi-table PDFs
RL with rewards for edit-distance, tables and formulas

the model improved dramatically on end to end across many categories, check the example below

weights: https://t.co/n0bxA3xrOo
code: https://t.co/cfjTam0bVd
paper: https://t.co/dWvK5t7eSV

please try it out and let us know your feedback. congrats to @griffintaur Wamiq Para @ChaySofian for leading this work

[Yasser Dahou @dahou_yasser]

We are releasing Falcon Perception, an open-vocabulary referring expression segmentation model. Along with it, a 0.3B OCR model that is on par with 3-10x larger competitors.

Current systems solve this with complex pipelines (separate encoders, late fusion, matching algorithms). We developed a novel simpler "bitter" approach: one early-fusion Transformer (image + text from first layer) with a shared parameter space, and let scale + training signal do the work. Please check our work !

📄 Paper: https://t.co/dWvK5t7MIt
💻 Code: https://t.co/AJ65GbMrUY
🎮 Playground: https://t.co/BIgisZkeid
🤗 Blogpost: https://t.co/J2IjlBPywF

[Yasser Dahou @dahou_yasser]

People can find all model variants here

https://t.co/IFuQi3pF8Q

We’ve added dense variants alongside the Agglomerative MoE models.

[Yasser Dahou @dahou_yasser]

Our Visual Reasoning Benchmark has been accepted to #CVPR2026

We wanted to know if VLMs can actually reason visually or if they're relying on text shortcuts.

well -> take away the text context, and even the best models struggle hard

we built a benchmark with 19k real images across 3 levels of difficulty.
level 1: basic perception. can you complete the pattern or fix the occlusion?
level 2: single rules. think raven's matrices but with real objects (color, count, orientation).
level 3: multi-attribute. complex rules mixing everything together.

paper: https://t.co/ToTZL2TAxE
hf page: https://t.co/e4Opz3ZpqF

with the amazing team
@BrigiMala @andyhuynh1111 @NarayanSanath @lkhphuc @ChaySofian @griffintaur