Video summary
This Tiny 0.8B AI Reads Documents Better Than OCR?
Main summary
Key takeaways
Summary
The video reviews LightOnOCR-3, a visual-language document-reading model, against Tesseract 5.4 across five document types. The 0.8B-parameter model was tested locally on an 8GB RTX 4060 laptop.
What LightOnOCR-3 does
- Converts page images into structured output, including Markdown, HTML tables, and LaTeX for equations.
- Its optional grounding mode labels page regions—such as text, titles, tables, graphs, and formulas—and provides their coordinates.
- According to the presenter’s reading of the model card, the supported prompts are empty or “grounding.” It is a document-reading tool, not a general chat model.
- The model is released under the Apache 2.0 license.
The presenter cites developer-reported OLMoOCR-Bench results of 85.5 for the 0.8B model and 86.3 for the 4B model. A larger model listed on the benchmark scores higher. The presenter cautions that these figures are developer-reported and may be affected by text-normalization choices.
Tests and results
Document Tesseract LightOnOCR-3 Note styled as handwriting Misread “six” as “at” and missed some characters. The presenter notes that the sample used a font, not actual handwriting. Correctly read the listed items and details, including “six eggs,” “6:00 PM,” and “room 204.” Table Recognized the numbers but returned them as lines of text, losing the columns. Produced an HTML table with all 15 numbers and the correct total. Grounding also identified the table region. Invoice Captured much of the text but disrupted the order and relationships between product names and amounts. Preserved the recipient block and product rows, and correctly extracted the stated charges and account details. Two-column research paper Did well with text order but mishandled an equation. Returned the equation in LaTeX and formatted headings as Markdown headings. Bar chart Found the title and some labels but not the chart’s full data. Reconstructed the values as a table: A 62, B 71, C 78, D 85.Speed and practical limits
On the test laptop, Tesseract took under 1.5 seconds per page. LightOnOCR-3 took 19 seconds for the note, 131 seconds for the table, 454 seconds for the invoice, and about 1,960 seconds for the research-paper page. The presenter attributes the slow local performance partly to some model operations falling back to regular PyTorch because optimized kernels were unavailable.
The blog’s reported 4.78 pages per second was achieved on an H100 using vLLM, a setup the presenter could not reproduce on their laptop. They also could not test the 4B model because it did not fit in 8GB of VRAM in BF16.
Takeaway: LightOnOCR-3’s advantage is preserving document structure and extracting data from tables, invoices, equations, and charts. For fast, straightforward text extraction from clean scans—especially on CPU—the presenter recommends Tesseract. The 0.8B model loaded in 5.5 seconds and used about 1.71GB of VRAM, but the tested Windows laptop was far too slow for practical page-by-page use.
Main speaker and sources
- Speaker: The video’s presenter from Prompt Engineer 48, reporting on their own tests.
- Sources discussed: LightOnAI’s model card and blog, the OLMoOCR-Bench results cited in the video, and Tesseract 5.4.
Rate this summary
Your feedback will help improve summaries.
Improve this summary
Reprocess with a stronger model when the summary feels incomplete or inaccurate.
Translate summary in another language
Ask questions to this video
Chat for follow-up questions, clarifications, and source-backed answers.