We Benchmarked 7 AI Models on Tamil OCR — Here Are the Real Numbers
7 AI மாடல்களை தமிழ் OCR-இல் சோதித்தோம் - உண்மையான முடிவுகள்
Tamil ChatGPT Team
Official Tamil ChatGPT Blog
We tested 7 vision-language models on clean, aged, and damaged Tamil document images and measured character-level accuracy against a known ground truth. One model hit 100% on all three at a fraction of the cost — here is the full methodology and data.
Most "Tamil OCR accuracy" claims online — including, previously, our own — are a bare percentage with no stated methodology: no sample images, no ground truth, no model name. We wanted to do better, so we built a small, transparent benchmark and are publishing the full results here.
Methodology
We created three test images of the same five-line Tamil paragraph, each representing a different real-world document condition:
- Clean — clear, printed Tamil text with no degradation.
- Aged — soft blur, faded contrast, speckle noise, and a 2.2° skew, simulating an older scanned document.
- Hard — stronger blur/fade/noise, a 3.4° skew, and a faded red "நகல்" (copy) stamp overlapping one line — simulating a stamped court or land-record copy.
Each image was sent to 7 vision-capable AI models via OpenRouter with an identical, strict "transcribe exactly, do not correct or guess" prompt. Output was scored against the known ground truth text at both character and word level.
Ground truth text
தமிழ் OCR சோதனை
தமிழ் மொழி உலகின் மிகப் பழமையான மொழிகளில் ஒன்றாகும்.
வணக்கம், எப்படி இருக்கிறீர்கள்?
தமிழ் ChatGPT சிறந்த AI கருவிகளை வழங்குகிறது.
இது ஒரு சோதனை ஆவணம்.
Results
| Model | Clean | Aged | Hard | Cost / page |
|---|---|---|---|---|
| deepseek/deepseek-v4.1-flash | 100% | 100% | 100% | $0.000074 (~$0.07 / 1,000 pages) |
| qwen/qwen3.7-flash | 100% | 100% | 100% | $0.000271 |
| z-ai/glm-5.3-flash | 100% | 100% | 100% | $0.000071 (slow: 16s on hard) |
| google/gemini-3.8-flash | 100% | 100% | 100% | $0.008232 (110× more) |
| inclusionai/ling-3.0-flash-vl | 100% | 98.6% | 91.7% | $0.000128 |
| google/gemini-2.5-flash-lite | 100% | 100% | 77.9% | $0.000210 |
| openai/gpt-5-nano | 0% (returned nothing) | — | — | — |
Winner: deepseek/deepseek-v4.1-flash — perfect character accuracy on all three samples, roughly $0.07 per 1,000 pages, 1-2 seconds per page. This is the model TamilChatGPT now uses for Tamil OCR, with automatic fallback to Gemini if OpenRouter is unavailable.
The "hard" sample is where most models broke down — gemini-2.5-flash-lite dropped to 77.9% character accuracy, truncating text where the "நகல்" stamp overlapped a line, and the smaller vision model showed similar failure patterns despite explicit instructions to transcribe exactly what was visible.
What this does and doesn't tell you
We want to be direct about the limits of this benchmark, not just the result:
- This tested 3 synthetic images of printed Tamil text, not a large corpus of real-world scanned documents. It is a real, repeatable test — not a statistically representative sample of all Tamil documents.
- It covers printed text only. We have not run an equivalent benchmark on handwritten Tamil, which is a harder problem and typically less accurate for any OCR system, including ours — our product pages say so explicitly.
- Results will shift as model providers ship new versions. We re-run this benchmark whenever a new vision model generation is released, and we will update this post when we do.
The full benchmark harness, sample images, and raw model outputs are available on request — we built it to be reproducible, not just quotable.
Want to try it yourself? Extract text from your own Tamil images with our OCR tool — free to start, no signup required.