How to Choose an OCR in 2026: Testing Nine Models on Three Inference Engines on Russian Handwriting

The article from Habr's AI hub compares a range of OCR solutions for Russian handwriting, testing nine models on vLLM, SGLang, and Native HF Transformers engines against a dataset of Russian handwritten texts. The authors report that Tesseract, despite being fastest at 193 pages per minute, returns only 411 characters per page, whereas PaddleOCR-VL yields 1193 tokens, indicating that higher speed often means missing content. The specialized PaddleOCR-VL (1.7B) outperforms the general-purpose Qwen2.5-VL-7B by 5.2 times in pages per minute and 6.7 times in tokens per second. Native HF Transformers lag behind vLLM and SGLang by two times for the same model due to lower GPU utilization. The benchmarks show that for simple printed documents, CPU-based Tesseract or RapidOCR are sufficient, while for complex layouts, tables, and Markdown output, PaddleOCR-VL with SGLang is recommended, and for production with minimal compatibility risks, PaddleOCR-VL with vLLM. For general-purpose tasks, Qwen2.5-VL-7B with vLLM is suitable, and for local RU-domain VLM, Cotype Light 3 with vLLM and MTP is suggested. The tests were conducted on a shared NVIDIA A100 80GB with batch=1 and 15 Russian handwriting samples, with unified settings like bfloat16 precision, temperature 0.0, and specific prompting.