TII Releases 0.27B-Param Falcon OCR 1.5, Scores 82.4 on olmOCR-Bench
HildeKuehne · x · 2026-09-12
TII has released Falcon OCR 1.5, a tiny 0.27B-parameter OCR model that scores 82.4 on olmOCR-Bench and 93.2 on OmniDocBench, with much lower hallucination rates and big gains on tables, handwriting, receipts and old scans.
Key techniques:
- New SFT data covering handwriting, aged scans, receipts and complex multi-table PDFs
- RL with rewards for edit distance, tables and formulas
- Dramatic end-to-end improvements across many categories
Weights, code and paper are all open-sourced; the team is soliciting community feedback.
Related event: TII Releases Falcon OCR 1.5, a 0.27B-Parameter OCR Model(2 posts)→
More from Models
- DeepSeek v4.1-Flash: 763B causal encoder-decoder with vision, KV cache cut to 1/8 — Latent Space · 2026-09-12
- Back-of-envelope: $200 Claude Max at full usage equals $2k-$4k of API value, implying negative margins — AaronBergman18 · 2026-09-12
- GPT-6 Astra Rebuilds Office Scan in Blender, Drops Humanoid Gym Into Newton Sim — liuziwei7 · 2026-09-12
- OpenAI reportedly ships Images 2.5, GPT-Live-1, Agents API ahead of DevDay — TheMoonMidas · 2026-09-12
- DeepSeek Back on Top With 6.6T Tokens Processed in a Single Day — gaganghotra_ · 2026-09-12
- The benchmarking problem: we're evaluating adapter-harness-model triplets, not models — teortaxesTex · 2026-09-12