ko-hand-ocr-vl โ€” ํ•œ๊ธ€ ์†๊ธ€์”จ OCR ์ „์šฉ VL ๋ชจ๋ธ (0.8B)

์‚ฌ์ง„์„ ๊ทธ๋Œ€๋กœ ๋„ฃ์œผ๋ฉด ์†๊ธ€์”จ๋ฅผ ๊ธ€์ž๋กœ ๋Œ๋ ค์ฃผ๋Š” ์˜คํ”ˆ ์›จ์ดํŠธ ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค. Qwen3.5-0.8B ๋ฅผ ํ•œ๊ธ€ ์†๊ธ€์”จ๋กœ ๋ฏธ์„ธ์กฐ์ •ํ–ˆ๊ณ  ๊ตฌ์กฐ๋Š” ํ•œ ์นธ๋„ ์•ˆ ๋ฐ”๊ฟจ์Šต๋‹ˆ๋‹ค. ๊ทธ๋ž˜์„œ Qwen3.5 ๋ฅผ ์˜ฌ๋ฆฌ๋Š” ๊ณณ โ€” Ollama, LM Studio, llama.cpp, vLLM, transformers โ€” ์— ์ƒˆ ์ฝ”๋“œ ์—†์ด ๊ทธ๋Œ€๋กœ ์˜ฌ๋ผ๊ฐ‘๋‹ˆ๋‹ค.

GGUF(LM Studioยทllama.cppยทOllama)๋Š” localdeel/ko-hand-ocr-vl-GGUF ์— ์žˆ์Šต๋‹ˆ๋‹ค.

๊ฐ™์€ ์ €์žฅ์†Œ ๊ณ„์—ด์˜ ko-hand-ocr(41M)๋Š” ์ž˜๋ผ ๋‘” ํ•œ ์ค„๋งŒ ์ฝ๋Š” ์ž‘๊ณ  ๋น ๋ฅธ ๋ชจ๋ธ์ด๊ณ , ์ด๊ฒƒ์€ ์‚ฌ์ง„ ํ†ต์งธ๋ฅผ ์ฝ์Šต๋‹ˆ๋‹ค. ์ž˜๋ผ ๋‘” ์ค„ ํ•˜๋‚˜๋งŒ ์ฝ์„ ๋•Œ๋Š” ๋‘˜์ด ๋น„์Šทํ•ฉ๋‹ˆ๋‹ค โ€” ์‹œํ—˜ ๊ธ€์”จ์ฒด๋Š” ์ด๊ฒƒ์ด, ์‚ฌ์ง„ ์ค„ ์นธ์€ 41M ์ด ์กฐ๊ธˆ ๋†’์Šต๋‹ˆ๋‹ค. 41M ์€ ๋ฐ›๋Š” ํฌ๊ธฐ๊ฐ€ 17๋ถ„์˜ 1์ด๊ณ  GPU ์—์„œ ์ค„์„ 3.9๋ฐฐ ๋นจ๋ฆฌ ์ฝ์Šต๋‹ˆ๋‹ค(์•„๋ž˜ ํ‘œ, ๊ฐ™์€ ํŒ์—์„œ ์žฐ ๊ฐ’).

2026-10-12 โ€” ์–ดํœ˜๋ฅผ ์ž˜๋ผ ๊ฐ€๋ฒผ์›Œ์กŒ์Šต๋‹ˆ๋‹ค. ์›๋ณธ Qwen3.5 ์˜ ์–ดํœ˜ 248,077๊ฐœ ๊ฐ€์šด๋ฐ ํ•œ๊ธ€ยท์ž๋ชจยท์ˆซ์žยท๊ธฐํ˜ธยท์งง์€ ์˜๋ฌธ ์กฐ๊ฐ๊ณผ ํ•™์Šตยท์‹œํ—˜์— ์“ฐ์ธ ํ† ํฐ๋งŒ ๋‚จ๊ฒจ 59,453๊ฐœ๋กœ ์ค„์˜€์Šต๋‹ˆ๋‹ค(ํŒŒ๋ผ๋ฏธํ„ฐ 853M -> 660M). ๋‹ค์‹œ ํ•™์Šตํ•˜์ง€ ์•Š์€ ๊ฐ™์€ ํŒ์ด๊ณ , ์‹œํ—˜ ๊ธ€์”จ์ฒดยท์ฒ˜์Œ ๋ณด๋Š” ๊ธ€์”จ์ฒดยท์‚ฌ์ง„ ์ค„ ์นธยท์‚ฌ์ง„ ํ†ต์งธ์—์„œ ๋‹ต์ด ๊ธ€์ž ํ•˜๋‚˜๊นŒ์ง€ ๊ฐ™์•˜์Šต๋‹ˆ๋‹ค โ€” transformers ๋กœ๋„, llama-server (Q8_0, ํ•œ ์ค„์”ฉ) ๋กœ๋„. ๋ฐ›๋Š” ํฌ๊ธฐ๋Š” safetensors 1.63GB -> 1.27GB, Q8_0 774MB -> 570MB, Q4_K_M 505MB -> 345MB ์ด๊ณ , llama.cpp ์—์„œ ๊ธ€์ž๋ฅผ ๋‚ด๋Š” ๋น ๋ฅด๊ธฐ๊ฐ€ 26~36% ์˜ฌ๋ž์Šต๋‹ˆ๋‹ค(์ด PC, llama-bench). ์–ดํœ˜์—์„œ ๋บ€ ๊ธด ์˜์–ด ๋‚ฑ๋ง๋กœ๋งŒ ์ง€์€ ์ค„ 200๊ฐœ์—์„œ๋Š” 173๊ฐœ๊ฐ€ ๊ฐ™์•˜๊ณ  ์ž๋ชจ ๋‹ฎ์Œ์ด 98.34 -> 98.01 ๋กœ ์กฐ๊ธˆ ๋‚ด๋ ค๊ฐ”์Šต๋‹ˆ๋‹ค. ์ค„ ๊ฒฌ์คŒํ‘œ์™€ ๋น ๋ฅด๊ธฐ๋Š” ์ž๋ฅธ ํŒ์œผ๋กœ ๋‹ค์‹œ ์žฐ ๊ฐ’์ด๊ณ , ์‚ฌ์ง„ ํ†ต์งธ์™€ ใ€Œ์–ด๋””์„œ ๋Œ๋ ค๋„ ๊ฐ™์€๊ฐ€ใ€ ํ‘œ๋Š” ์ž๋ฅด๊ธฐ ์ „์— ์žฐ ๊ฐ’์ž…๋‹ˆ๋‹ค(๋‹ต์ด ๊ฐ™์•„ ๊ทธ๋Œ€๋กœ ๋‘ก๋‹ˆ๋‹ค). ์ž๋ฅด๊ธฐ ์ „ ํŒ์€ ์ด ์ €์žฅ์†Œ์˜ ํƒœ๊ทธ full-vocab ์— ์žˆ์Šต๋‹ˆ๋‹ค.

์–ผ๋งˆ๋‚˜ ์ฝ๋‚˜

๋ชจ๋‘ ๊ฐ™์€ ์‹œํ—˜์ง€, ๊ฐ™์€ PC(RTX 5060 Laptop 8GB)์—์„œ ์žฐ ๊ฐ’์ž…๋‹ˆ๋‹ค. ์‹œํ—˜ ๊ธ€์”จ์ฒด๋Š” ํ•™์Šต์—์„œ ๋บ€ ๊ธ€๊ผด์ด๊ณ , ์‚ฌ์ง„ 11์žฅ์€ ์‚ฌ๋žŒ์ด ์‹ค์ œ๋กœ ์“ด ์†๊ธ€์”จ์ด๋ฉฐ ํ•™์Šต์— ํ•œ ์žฅ๋„ ์“ฐ์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค(๊ณต๊ฐœํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค). ์ž๋ชจ ๋‹ฎ์Œ์€ ๋†’์„์ˆ˜๋ก, CER(๊ธ€์ž ์˜ค๋ฅ˜์œจ)์€ ๋‚ฎ์„์ˆ˜๋ก ์ข‹์Šต๋‹ˆ๋‹ค.

์ค„ ํ•˜๋‚˜์”ฉ โ€” ๋‹ค๋ฅธ OCR ๊ณผ ๊ฐ™์€ ์ž๋กœ

๊ฐ™์€ ์‹œํ—˜์ง€๋กœ ์—ด ๋ชจ๋ธ์„ ๊ฒฌ์ค€ ์ž๋ชจ ๋‹ฎ์Œ

๋ชจ๋ธ ํฌ๊ธฐ ์‹œํ—˜ ๊ธ€์”จ์ฒด 6๋ฒŒ (์ž๋ชจ ๋‹ฎ์Œ / CER) ์ฒ˜์Œ ๋ณด๋Š” ๊ธ€์”จ์ฒด 24๋ฒŒ ์‚ฌ์ง„ 11์žฅ ์ค„ ์นธ (์ž๋ชจ ๋‹ฎ์Œ / CER) GPU ์ค„/์ดˆ
ko-hand-ocr-vl (์ด๊ฒƒ, 660M) 1.27GB 97.0 / 5.8 97.4 93.4 / 15.8 10.8
Qwen3.5-0.8B (๋ฏธ์„ธ์กฐ์ • ์ „ ์›๋ณธ) 1.63GB 53.0 / 161.8 51.4 54.2 / 115.3 9.6
ko-hand-ocr ๋‹จ์ผ ๋ชจ๋ธ (41M, ์ค„ ์ „์šฉ) 78MB 96.8 / 7.2 97.3 94.2 / 13.4 42.0
ko-hand-ocr ์•™์ƒ๋ธ” (๋ชจ๋ธ 3๊ฐœ) 225MB 97.1 / 6.7 97.7 96.5 / 8.6 8.3
Qwen/Qwen3-VL-2B-Instruct 3.96GB 66.6 / 79.1 68.0 69.9 / 38.3 7.3
PaddlePaddle/PaddleOCR-VL-1.6 1.79GB 63.6 / 62.9 66.5 61.7 / 44.0 1.2
ddobokki/ko-trocr 408MB 78.2 / 42.3 80.4 78.5 / 39.2 6.1
PaddleOCR PP-OCRv5 (korean) 13MB 72.7 / 40.5 75.3 71.8 / 39.7 397.4
EasyOCR (korean) 15MB 56.6 / 64.3 64.6 66.9 / 55.0 81.3
Tesseract 5 (kor+eng) 6MB 17.7 / 97.3 19.7 13.2 / 98.6 โ€”

2026-10-13 ์— ์—ด ๋ชจ๋ธ์„ ๊ฐ™์€ PC ์—์„œ ํ•œ ๋ฒˆ์— ์Ÿ€์Šต๋‹ˆ๋‹ค. ์‹œํ—˜ ๊ธ€์”จ์ฒด๋Š” 6๋ฒŒ ํ•ฉ์ณ 640์ค„, ์ค„ ์นธ์€ ko-hand-ocr ์˜ ์ค„ ์ž๋ฅด๊ธฐ๋กœ ์ž˜๋ผ ๋ชจ๋‘์—๊ฒŒ ๊ฐ™์€ ์นธ์„ ์ค๋‹ˆ๋‹ค. ์ƒ๋Œ€์—๊ฒŒ๋Š” ๋ชจ๋ธ ์นด๋“œ์˜ ์‚ฌ์šฉ๋ฒ• ๊ทธ๋Œ€๋กœ, ์š•์‹ฌ ๋ฝ‘๊ธฐ(greedy)๋กœ ๋ฌป์Šต๋‹ˆ๋‹ค. ๋ฌผ์Œ์€ Qwen ๊ณ„์—ด์ด ใ€Œ์ด ์ด๋ฏธ์ง€์— ์“ฐ์ธ ๊ธ€์ž๋ฅผ ๊ทธ๋Œ€๋กœ ์ฝ์–ด ์ ์œผ์„ธ์š”. ๊ธ€์ž๋งŒ ์ถœ๋ ฅํ•˜์„ธ์š”.ใ€, PaddleOCR-VL ์ด ใ€ŒOCR:ใ€ ์ž…๋‹ˆ๋‹ค.

์‚ฌ์ง„ ํ†ต์งธ

์‚ฌ์ง„ ํ•œ ์žฅ์„ ์›๋ณธ ๊ทธ๋Œ€๋กœ ์ค๋‹ˆ๋‹ค. ๊ทธ๋ฆผ์€ 4,096ํ† ํฐ๊นŒ์ง€๋กœ ์ค„์—ฌ ๋“ค์–ด๊ฐ‘๋‹ˆ๋‹ค(llama.cppยทOllama ์™€ ๊ฐ™์€ ์ƒํ•œ). ์ชฝ ์€ ์ชฝ์˜ ๊ธ€์ž๋ฅผ ๋‹ค ์ด์–ด ๋ถ™์—ฌ ๊ฒฌ์ค€ ๊ฐ’(ํ†ต์งธ OCR ์„ ์žฌ๋Š” ํ”ํ•œ ์ž)์ด๊ณ , ์ค„ ์ง ์€ ์ฝ์€ ์ค„์„ ์ •๋‹ต ์ค„๊ณผ ์ง์ง€์–ด ์žฐ ๊ฐ’์ž…๋‹ˆ๋‹ค. ์ •๋‹ตํ‘œ๊ฐ€ ํ•œ ์ค„์— ๋‚˜๋ž€ํžˆ ์“ด ใ€Œ๋ถ€์„œ: โ€ฆ ์„ฑ๋ช…: โ€ฆใ€ ์„ ๋‘ ์ค„๋กœ ๋‚˜๋ˆ  ์ ์–ด ๋‘์–ด์„œ, ํ•œ ์ค„๋กœ ๋ฐ”๋ฅด๊ฒŒ ์ฝ์–ด๋„ ์ค„ ์ง์œผ๋กœ๋Š” ๊นŽ์ž…๋‹ˆ๋‹ค. ์ƒ๋Œ€ VL ๋‘˜์€ ์šฐ๋ฆฌ ์ƒํ•œ์„ ์”Œ์šฐ์ง€ ์•Š๊ณ  ๊ทธ์ชฝ ๊ธฐ๋ณธ ๊ทธ๋ฆผ ํฌ๊ธฐ ๊ทธ๋Œ€๋กœ ์ค๋‹ˆ๋‹ค.

์‚ฌ์ง„ ํ†ต์งธ โ€” ์ชฝ ๋‹จ์œ„ ๋‹ฎ์Œ๊ณผ ์˜ค๋ฅ˜์œจ

๋ชจ๋ธ ์ชฝ ์ž๋ชจ ๋‹ฎ์Œ ์ชฝ CER ๊ฐ€์žฅ ๋‚ฎ์€ ์‚ฌ์ง„(์ชฝ) ์ค„ ์ง ์ž๋ชจ ๋‹ฎ์Œ
ko-hand-ocr-vl (์ด๊ฒƒ) 94.3 15.3 78.3 95.6
Qwen3.5-0.8B (์›๋ณธ) 72.7 42.1 55.3 62.1
Qwen3-VL-2B-Instruct 76.9 41.1 47.6 72.5
PaddleOCR-VL-1.6 79.9 37.3 54.5 57.3

์–ด๋””์„œ ๋Œ๋ ค๋„ ๊ฐ™์€๊ฐ€

๊ฐ™์€ ๋ชจ๋ธ์„ ๋ฐ›๋Š” ๊ธธ๋งˆ๋‹ค ๋‹ค์‹œ ์Ÿ€์Šต๋‹ˆ๋‹ค. ์–‘์žํ™”๋‚˜ ๊ทธ๋ฆผ ํฌ๊ธฐ ๋งž์ถ”๊ธฐ๊ฐ€ ๊ธธ๋งˆ๋‹ค ๋‹ฌ๋ผ์„œ ๊ทธ๋Œ€๋กœ ๋ฏฟ์ง€ ์•Š๊ณ  ์Ÿ€์Šต๋‹ˆ๋‹ค(์ด ํ‘œ๋Š” ๊ธ€์”จ์ฒด ๋ฒŒ๋งˆ๋‹ค 100์ค„). llama.cpp ์ชฝ(LM StudioยทOllama ํฌํ•จ)์ด ์‚ฌ์ง„์—์„œ ๋‚ฎ์€ ๊ฒƒ์€ ์–‘์žํ™” ํƒ“์ด ์•„๋‹™๋‹ˆ๋‹ค โ€” ์–‘์žํ™” ์•ˆ ํ•œ F16 GGUF ๋„ ๊ฐ™์€ ๊ฐ’์ด ๋‚˜์˜ต๋‹ˆ๋‹ค. ๊ทธ๋ฆผ ํฌ๊ธฐ๋ฅผ ๋งž์ถ”๋Š” ๋ฒ•์ด transformers ์™€ ๋‹ฌ๋ผ์„œ์ž…๋‹ˆ๋‹ค. llama.cpp ๋Š” ๊ฐ€๋กœ์„ธ๋กœ ๋น„๋ฅผ ์ง€ํ‚ค๋ฉฐ ์ค„์ด๊ณ  ๋‚จ๋Š” ๊ฐ€์žฅ์ž๋ฆฌ๋ฅผ ๊ฒ€์ •์œผ๋กœ ์ฑ„์šฐ๋ฉฐ, ์ž‘์€ ๊ทธ๋ฆผ์„ 8ํ† ํฐ๊นŒ์ง€ ์ค„์ž…๋‹ˆ๋‹ค(์ด ๋ชจ๋ธ์€ 64ํ† ํฐ๋ถ€ํ„ฐ ๋ฐฐ์› ์Šต๋‹ˆ๋‹ค).

๊ทธ๋ž˜์„œ ๊ทธ๋ฆผ ์ตœ์†Œยท์ตœ๋Œ€ ํฌ๊ธฐ(64~4,096ํ† ํฐ)๋ฅผ mmproj ํŒŒ์ผ ์•ˆ์— ์ ์–ด ๋‘์—ˆ์Šต๋‹ˆ๋‹ค(2026-10-09). llama.cpp ์™€ LM Studio ๋Š” ๋ฐ›๊ธฐ๋งŒ ํ•˜๋ฉด ์ด ๊ฐ’์„ ์”๋‹ˆ๋‹ค โ€” ๊ฐ™์€ LM Studio ์—์„œ ์‚ฌ์ง„ ์ค„ ์นธ์ด 90.8 -> 92.5, llama-server ์—์„œ 90.9 -> 92.0 ์œผ๋กœ ์˜ฌ๋ž์Šต๋‹ˆ๋‹ค. Ollama 0.34.4 ๋Š” ์•„์ง ์ด ๊ฐ’์„ ์ฝ์ง€ ์•Š์•„ ๊ทธ๋Œ€๋กœ์ž…๋‹ˆ๋‹ค. ๊ฒ€์ • ์—ฌ๋ฐฑ์€ ํŒŒ์ผ๋กœ ๋ฐ”๊ฟ€ ์ˆ˜ ์—†์Šต๋‹ˆ๋‹ค.

๋Œ๋ฆฌ๋Š” ๊ธธ ์‹œํ—˜ ๊ธ€์”จ์ฒด 6๋ฒŒ ์‚ฌ์ง„ ์ค„ ์นธ ์‚ฌ์ง„ ํ†ต์งธ
transformers (bf16, safetensors) 96.9 93.3 94.3
llama.cpp llama-server โ€” Q8_0 96.7 92.0 91.8
LM Studio โ€” Q8_0, ๋ฐ›์€ ๊ทธ๋Œ€๋กœ (์„ค์ • ์•ˆ ๊ฑด๋“œ๋ฆผ) 96.4 92.5 91.8
Ollama โ€” hf.co/โ€ฆ:Q8_0, ๋ฐ›์€ ๊ทธ๋Œ€๋กœ 96.2 90.9 91.8

์–ด๋–ค ์ค„์„ ํ‹€๋ฆฌ๋‚˜

์ค„ ์•ˆ์— ๋ฌด์—‡์ด ๋“ค์—ˆ๋Š”์ง€์— ๋”ฐ๋ผ ๋‚˜๋ˆˆ ๊ธ€์ž ์˜ค๋ฅ˜์œจ(CER)์ž…๋‹ˆ๋‹ค. ๋‚ฎ์„์ˆ˜๋ก ์ข‹์Šต๋‹ˆ๋‹ค.

๊ฐˆ๋ž˜๋ณ„ ๊ธ€์ž ์˜ค๋ฅ˜์œจ

๋น ๋ฅด๊ธฐ์™€ ๋ฉ์น˜

๋น ๋ฅด๊ธฐ์™€ ๋ฉ์น˜

๋ฒค์น˜๋งˆํฌ๋Š” ์ด๋ ‡๊ฒŒ ์Ÿ€์Šต๋‹ˆ๋‹ค

์œ„ ์ˆซ์ž๊ฐ€ ์–ด๋””์„œ ๋‚˜์™”๋Š”์ง€ ๋‹ค ์ ์Šต๋‹ˆ๋‹ค. ๊ฐ™์€ ์‹œํ—˜์ง€๋ฅผ ๊ฐ™์€ PC ์—์„œ, ์—ด ๋ชจ๋ธ์„ ํ•œ ๋ฒˆ์— ๋Œ๋ ธ๊ณ (2026-10-13 14:59) ์‚ฌ์ง„์€ ํ•™์Šต์— ํ•œ ์žฅ๋„ ์“ฐ์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค.

์‹œํ—˜์ง€

์‹œํ—˜์ง€ ๋ฌด์—‡์ธ๊ฐ€ ํฌ๊ธฐ ํ•™์Šต์— ์ผ๋‚˜
์‹œํ—˜ ๊ธ€์”จ์ฒด 6๋ฒŒ ๋ฌด๋ฃŒ ์†๊ธ€์”จ ๊ธ€๊ผด ์—ฌ์„ฏ ๋ฒŒ(OFL)๋กœ ์ค„์„ ๊ทธ๋ฆฐ ๊ทธ๋ฆผ. ํ•™์Šต ๊ทธ๋ฆผ๊ณผ ๊ฐ™์€ ๊ทธ๋ฆฌ๊ฐœ๋กœ ์ข…์ด ๊ฒฐยท๊ธฐ์šธ๊ธฐยท์ง€์šด ์ž๊ตญ ๊ฐ™์€ ํ”๋“ค๊ธฐ๋ฅผ ์ž…ํ˜”๊ณ , ์”จ์•—์€ ํ•™์Šต์— ์•ˆ ์“ฐ๋Š” ๊ณ ์ •๊ฐ’์ž…๋‹ˆ๋‹ค 640์ค„ 4,448์ž ์•ˆ ์”€. ํ•™์Šต ๊ธ€๊ผด 450๋ฒŒ ๊ฐ€์šด๋ฐ ์ด ์—ฌ์„ฏ๊ณผ ์ƒ๊น€์ƒˆ๊ฐ€ ๋‹ฎ์€ ๊ฒƒ(๋‹ฎ์Œ 0.7 ์ด์ƒ)๋„ ๋บ์Šต๋‹ˆ๋‹ค
์ฒ˜์Œ ๋ณด๋Š” ๊ธ€์”จ์ฒด 24๋ฒŒ ๋‚˜๋ˆ”์†๊ธ€์”จ 6 ยท ๊ทธ ๋ฐ– 17 ยท ๊ตฌ๊ธ€ 1 (์•„๋ž˜ ๋ชฉ๋ก). ๊ฐ™์€ ๋ฐฉ์‹์œผ๋กœ ๊ทธ๋ฆผ 634์ค„ 3,614์ž ์•ˆ ์”€. ๋ชจ๋ธ์„ ๊ณ ๋ฅด๋Š” ๋ฐ๋„ ์•ˆ ์”€ โ€” ๊ทธ๋ž˜์„œ ใ€Œ๋‘๋ฃจ ์ฝ๋‚˜ใ€๋ฅผ ๋ฌป๋Š” ์ž์ž…๋‹ˆ๋‹ค
์‚ฌ์ง„ 11์žฅ โ€” ์ค„ ์นธ ์‚ฌ๋žŒ์ด ์‹ค์ œ ์„œ์‹ ์ข…์ด์— ์†์œผ๋กœ ์จ์„œ ์ฐ์€ ์‚ฌ์ง„. ko-hand-ocr ์˜ ์ค„ ์ž๋ฅด๊ธฐ๋กœ ์ž˜๋ผ ๋ชจ๋‘์—๊ฒŒ ๊ฐ™์€ ์นธ์„ ์ค๋‹ˆ๋‹ค 34์ค„ 209์ž ์•ˆ ์”€. ์‚ฌ๋žŒ ์ž๋ฃŒ๋ผ ๊ณต๊ฐœํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค
์‚ฌ์ง„ 11์žฅ โ€” ํ†ต์งธ ๊ฐ™์€ ์‚ฌ์ง„์„ ์ž๋ฅด์ง€ ์•Š๊ณ  ํ•œ ์žฅ ๊ทธ๋Œ€๋กœ 11์žฅ ใ€ƒ

์‹œํ—˜ ๊ธ€์”จ์ฒด ์‹œํ—˜์ง€

์‹œํ—˜ ๊ธ€์”จ์ฒด ์ค„์€ ์„œ์‹์—์„œ ์‹ค์ œ๋กœ ๋ณด๋Š” ๊ผด์„ ์„ž์–ด ๋งŒ๋“ญ๋‹ˆ๋‹ค. ๋Œ€๋žต ์…‹์— ํ•˜๋‚˜๊ฐ€ ๋œป ์—†๋Š” ์Œ์ ˆ์„ ์ด์–ด ๋ถ™์ธ ์ค„(์ด๋ฆ„ยท์ƒํ˜ธ์ฒ˜๋Ÿผ ์–ธ์–ด ์ง€์‹์ด ์•ˆ ํ†ตํ•˜๋Š” ์ค„ โ€” ํš๋งŒ ๋ณด๊ณ  ์ฝ๋Š” ํž˜์„ ์žฝ๋‹ˆ๋‹ค), ์…‹์— ํ•˜๋‚˜๊ฐ€ ์„œ์‹ ์นธ(ใ€Œ์„ฑ๋ช… : โ€ฆใ€, ใ€Œ์ผ์ž : 2026.03.17ใ€), ๋‹ค์„ฏ์— ํ•˜๋‚˜๊ฐ€ ์ œ๋ชฉยท๋ผ๋ฒจ, ๋‚˜๋จธ์ง€๊ฐ€ ๋ฌธ์žฅ์ž…๋‹ˆ๋‹ค. ์ˆซ์žยท๋‚ ์งœยท์˜๋ฌธ ์•ฝ์–ดยท๊ธฐํ˜ธ๊ฐ€ ์„ž์ž…๋‹ˆ๋‹ค. ํ™‘์ž๋ชจ ์ค„๊ณผ ๊ทธ ๊ธ€๊ผด์ด ๋ชป ๊ทธ๋ฆฌ๋Š” ๊ธ€์ž๊ฐ€ ๋“  ์ค„์€ ๋บ๋‹ˆ๋‹ค โ€” ์•„๋ฌด๋„ ๋งžํž ์ˆ˜ ์—†๋Š” ์ค„์ด๋ผ์„œ์ž…๋‹ˆ๋‹ค(2026-10-08 ์— ๊ณ ์นœ ์ž. ๊ทธ ์ „ ์ˆซ์ž์™€ ์„ž์–ด ๊ฒฌ์ฃผ์ง€ ์•Š์Šต๋‹ˆ๋‹ค).

์‚ฌ์ง„ ์ชฝ 12๋ฒˆ์งธ ์žฅ(์ €ํ•ด์ƒ๋„ ์ธ์‡„ ์„œ์‹ + ์„œ๋ช…)์€ ์†๊ธ€์”จ ํŒ๋…์˜ ๋ชฉํ‘œ๊ฐ€ ์•„๋‹ˆ๋ผ์„œ ์žฌ๊ธฐ๋งŒ ํ•˜๊ณ  ์ ์ˆ˜์—์„œ๋Š” ๋บ๋‹ˆ๋‹ค. ์ด ๋ชจ๋ธ์€ ๊ทธ ์žฅ์—์„œ 61.3% ์ž…๋‹ˆ๋‹ค.

์ฒ˜์Œ ๋ณด๋Š” ๊ธ€์”จ์ฒด 24๋ฒŒ ๋ชฉ๋ก

๋‚˜๋ˆ”๋ธŒ๋Ÿฌ์‹œ์Šคํฌ๋ฆฝํŠธ, KCC ๋„๋‹ด๋„๋‹ด์ฒด, ๊ฐ€๋น„์•„ ๋ด„๋ฐ”๋žŒ์ฒด, ๊ทธ๋ฆฌ์šด ๊ทœ์›์ฒด, ๊ทธ๋ฆฌ์šด ํ”„๋กฌ์†”, ๋‚ ์Œ˜๋Œ์ด์ฒด, ์„ธ๊ตฌ์„ธ๊ตฌ์ฒด, ์–ด๋น„ ๊น€๋น ๋ฅธ์ฒด, ์–ด๋น„ ๋งˆ์ธ ์ฝ”์ฒด, ์–ด๋น„ ์Šค์œ—์ฒด, ์–ด๋น„ ์€๋””์ฒด, ์–ด๋น„ ํ€ธ์ œ์ด์ฒด, ์˜จ๊ธ€์žŽ ๊ณต๋ถ€์ž˜ํ•˜์ž๋‚˜, ์˜จ๊ธ€์žŽ ์–ธ์ฆˆ์ฒด, ์œค์ดˆ๋ก์šฐ์‚ฐ์–ด๋ฆฐ์ด ๋ฏผ๊ตญ, ์นดํŽ˜24 ๋™๋™, ํ•™๊ต์•ˆ์‹ฌ ๊พธ๋Ÿฌ๊ธฐ, ํ•™๊ต์•ˆ์‹ฌ ์น ํŒ์ง€์šฐ๊ฐœ, ๋‚˜๋ˆ”์†๊ธ€์”จ ๊ธฐ์จ๋ฐ์Œ, ๋‚˜๋ˆ”์†๊ธ€์”จ ๋”ฐ๋œปํ•œ ์ž‘๋ณ„, ๋‚˜๋ˆ”์†๊ธ€์”จ ๋ถ€์žฅ๋‹˜ ๋ˆˆ์น˜์ฒด, ๋‚˜๋ˆ”์†๊ธ€์”จ ์•„๋น ๊ธ€์”จ, ๋‚˜๋ˆ”์†๊ธ€์”จ ์™ธํ• ๋จธ๋‹ˆ๊ธ€์”จ, ๋‚˜๋ˆ”์†๊ธ€์”จ ํ•˜๋žŒ์ฒด

์ ์ˆ˜

์ ์ˆ˜ ์–ด๋–ป๊ฒŒ ์…ˆํ•˜๋‚˜
์ž๋ชจ ๋‹ฎ์Œ ์ฝ์€ ๊ฒƒ๊ณผ ์ •๋‹ต์„ ์ž๋ชจ(ใ„ฑ ใ… ใ… โ€ฆ)๋กœ ํ’€๊ณ  ๋„์–ด์“ฐ๊ธฐ๋ฅผ ์ง€์šด ๋’ค ์–ผ๋งˆ๋‚˜ ๋‹ฎ์•˜๋‚˜(ํŒŒ์ด์ฌ difflib ๋น„์œจ). ใ€ŒํŒ€ใ€์„ ใ€Œํƒใ€์œผ๋กœ ์ฝ์œผ๋ฉด ์ž๋ชจ ์…‹ ๊ฐ€์šด๋ฐ ํ•˜๋‚˜๋งŒ ํ‹€๋ ค 0.67 ์ž…๋‹ˆ๋‹ค. ์ค„๋งˆ๋‹ค ๋‚ด์„œ ํ‰๊ท ํ•ฉ๋‹ˆ๋‹ค
CER(๊ธ€์ž ์˜ค๋ฅ˜์œจ) ์ •๋‹ต์ด ๋˜๋„๋ก ๊ณ ์ณ์•ผ ํ•˜๋Š” ๊ธ€์ž ์ˆ˜(ํŽธ์ง‘ ๊ฑฐ๋ฆฌ) รท ์ •๋‹ต ๊ธ€์ž ์ˆ˜, ๋„์–ด์“ฐ๊ธฐ๋Š” ์ง€์šฐ๊ณ . ์ค„๋งˆ๋‹ค ํ‰๊ท  ๋‚ด์ง€ ์•Š๊ณ  ๊ธ€์ž๋ฅผ ๋‹ค ๋ชจ์•„ ๋‚˜๋ˆ•๋‹ˆ๋‹ค. ์ •๋‹ต๋ณด๋‹ค ๊ธธ๊ฒŒ ์ง€์–ด๋‚ด๋ฉด 100% ๋ฅผ ๋„˜์Šต๋‹ˆ๋‹ค(์›๋ณธ Qwen3.5 ์˜ 161.8% ๊ฐ€ ๊ทธ ๊ฒฝ์šฐ)
๊ฐ€์žฅ ๋‚ฎ์€ ๋ฒŒ / ์‚ฌ์ง„ ํ‰๊ท ์ด ๊ฐ€๋ฆฌ๋Š” ๋ฐ”๋‹ฅ. ์—ฌ์„ฏ ๋ฒŒ ์‚ฌ์ด๊ฐ€ ๋ช‡ ์ ์”ฉ ๋ฒŒ์–ด์ง€๊ธฐ ๋•Œ๋ฌธ์— ๋”ฐ๋กœ ์ ์Šต๋‹ˆ๋‹ค
์‚ฌ์ง„ ํ†ต์งธ โ€” ์ชฝ ์ฝ์€ ๊ธ€์ž๋ฅผ ์ชฝ ์ „์ฒด๋กœ ์ด์–ด ๋ถ™์—ฌ ์ •๋‹ต ์ „์ฒด์™€ ๊ฒฌ์ค๋‹ˆ๋‹ค(ํ†ต์งธ OCR ์„ ์žฌ๋Š” ํ”ํ•œ ์ž)
์‚ฌ์ง„ ํ†ต์งธ โ€” ์ค„ ์ง ์ฝ์€ ์ค„๊ณผ ์ •๋‹ต ์ค„์„ ๊ฐ€์žฅ ๋‹ฎ์€ ๊ฒƒ๋ผ๋ฆฌ ์ง์ง€์–ด ์žฝ๋‹ˆ๋‹ค. ์ •๋‹ตํ‘œ๊ฐ€ ํ•œ ์ค„์— ๋‚˜๋ž€ํžˆ ์“ด ๋‘ ์นธ์„ ๋‘ ์ค„๋กœ ์ ์–ด ๋‘์–ด์„œ, ํ•œ ์ค„๋กœ ๋ฐ”๋ฅด๊ฒŒ ์ฝ์–ด๋„ ๊นŽ์ž…๋‹ˆ๋‹ค

์žฅ๋น„์™€ ํŒ

GPU NVIDIA GeForce RTX 5060 Laptop GPU (8GB)
CPU AMD Ryzen 7 260 (๋…ธํŠธ๋ถ, 8์ฝ”์–ด 16์Šค๋ ˆ๋“œ)
์šด์˜์ฒด์ œ Windows 11 Pro
PyTorch 2.14.0+cu130
์ด ๋ชจ๋ธ (transformers) transformers 5.16.1, bfloat16, ์š•์‹ฌ ๋ฝ‘๊ธฐ(greedy), ์ค„ ํ•˜๋‚˜์— ์ตœ๋Œ€ 64ํ† ํฐ. ์‚ฌ์ง„ ํ†ต์งธ๋Š” ๊ทธ๋ฆผ์„ 4,096ํ† ํฐ๊นŒ์ง€๋กœ ์ค„์—ฌ ๋„ฃ์Œ
๋ฌผ์Œ ใ€Œ์ด ์ด๋ฏธ์ง€์— ์“ฐ์ธ ๊ธ€์ž๋ฅผ ๊ทธ๋Œ€๋กœ ์ฝ์–ด ์ ์œผ์„ธ์š”. ๊ธ€์ž๋งŒ ์ถœ๋ ฅํ•˜์„ธ์š”.ใ€
llama.cpp llama.cpp llama-server build 11474 (b9acf138a), CUDA
LM Studio LM Studio โ€” ์—”์ง„ llama.cpp CUDA 12 2.55.0, ๋ฐ›์€ ์„ค์ • ๊ทธ๋Œ€๋กœ
Ollama Ollama 0.34.4, ๋ฐ›์€ ์„ค์ • ๊ทธ๋Œ€๋กœ

๋ชจ๋ธ๋งˆ๋‹ค ์–ด๋–ป๊ฒŒ ๋Œ๋ ธ๋‚˜

์ƒ๋Œ€๋Š” ๋ชจ๋‘ ๋ชจ๋ธ ์นด๋“œ์˜ ์‚ฌ์šฉ๋ฒ• ๊ทธ๋Œ€๋กœ ๋Œ๋ ธ์Šต๋‹ˆ๋‹ค. ์—”์ง„๋งˆ๋‹ค ๋”ฐ๋กœ ๊น ํ™˜๊ฒฝ์—์„œ ๋Œ๋ฆฌ๊ณ , ์‹œ๊ฐ„์€ ์—”์ง„ ์•ˆ์—์„œ ์žฐ ๊ฐ’๋งŒ ์…‰๋‹ˆ๋‹ค(๊ทธ๋ฆผ์„ ๋„˜๊ธฐ๋Š” ๊ฐ’์€ ๋บ๋‹ˆ๋‹ค).

๋ชจ๋ธ ์–ด๋–ป๊ฒŒ
ko-hand-ocr ์•™์ƒ๋ธ” ๋ชจ๋ธ 3๊ฐœ x ํ”๋“ค๊ธฐ ์…‹์„ ์ฝ๊ณ  ์„œ๋กœ ๊ฐ€์žฅ ๋‹ฎ์€ ๋‹ต(๊ฐ€์šด๋ฐ ๋‹ต)
ko-hand-ocr ๋‹จ์ผ ๋ชจ๋ธ ko-hand-ocr 0.4.1 Reader, float32, ๋น” 5, ํ”๋“ค๊ธฐ ์…‹
ddobokki/ko-trocr transformers, float32, ๋น” 5, ๊ธธ์ด ์ƒํ•œ 16 -> 64 ๋กœ ํ’€์–ด ์คŒ
PaddleOCR PP-OCRv5 (korean) paddleocr TextRecognition(korean_PP-OCRv5_mobile_rec) (paddleocr 3.7.0 / paddle 3.4.0)
EasyOCR (korean) easyocr korean_g2 (ko+en), recognize() ์— ์ค„ ๊ทธ๋ฆผ ์ „์ฒด๋ฅผ ํ•œ ์นธ์œผ๋กœ (1.7.2)
Tesseract 5 (kor+eng) tesseract 5.5.3, --oem 1 --psm 7 -l kor+eng
PaddlePaddle/PaddleOCR-VL-1.6 transformers 5.16.1, bfloat16, ๋ฌผ์Œ 'OCR:', ์š•์‹ฌ ๋ฝ‘๊ธฐ, ์ตœ๋Œ€ 64ํ† ํฐ
Qwen/Qwen3-VL-2B-Instruct transformers 5.16.1, bfloat16, ๋ฌผ์Œ '์ด ์ด๋ฏธ์ง€์— ์“ฐ์ธ ๊ธ€์ž๋ฅผ ๊ทธ๋Œ€๋กœ ์ฝ์–ด ์ ์œผ์„ธ์š”. ๊ธ€์ž๋งŒ ์ถœ๋ ฅํ•˜์„ธ์š”.', ์š•์‹ฌ ๋ฝ‘๊ธฐ, ์ตœ๋Œ€ 64ํ† ํฐ
ko-hand-ocr-vl (Qwen3.5-0.8B ๋ฏธ์„ธ์กฐ์ •) transformers 5.16.1, bfloat16, ๋ฌผ์Œ '์ด ์ด๋ฏธ์ง€์— ์“ฐ์ธ ๊ธ€์ž๋ฅผ ๊ทธ๋Œ€๋กœ ์ฝ์–ด ์ ์œผ์„ธ์š”. ๊ธ€์ž๋งŒ ์ถœ๋ ฅํ•˜์„ธ์š”.', ์š•์‹ฌ ๋ฝ‘๊ธฐ, ์ตœ๋Œ€ 64ํ† ํฐ
Qwen/Qwen3.5-0.8B transformers 5.16.1, bfloat16, ๋ฌผ์Œ '์ด ์ด๋ฏธ์ง€์— ์“ฐ์ธ ๊ธ€์ž๋ฅผ ๊ทธ๋Œ€๋กœ ์ฝ์–ด ์ ์œผ์„ธ์š”. ๊ธ€์ž๋งŒ ์ถœ๋ ฅํ•˜์„ธ์š”.', ์š•์‹ฌ ๋ฝ‘๊ธฐ, ์ตœ๋Œ€ 64ํ† ํฐ

๊ธ€์”จ์ฒด๋งˆ๋‹ค

ํ•ฉ๊ณ„ ํ•˜๋‚˜๋งŒ ๋ณด๋ฉด ๋ฌด๋„ˆ์ง€๋Š” ๊ธ€์”จ์ฒด๊ฐ€ ๊ฐ€๋ ค์ง‘๋‹ˆ๋‹ค. ๋ฒŒ๋งˆ๋‹ค ์ค„ ์ˆ˜๊ฐ€ ๋‹ค๋ฅธ ๊ฒƒ์€ ๋ชป ๊ทธ๋ฆฌ๋Š” ์ค„์„ ๋บ๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค.

๊ธ€์”จ์ฒด ์ค„ ์ด ๋ชจ๋ธ 41M ๋‹จ์ผ ๋ชจ๋ธ 41M ์•™์ƒ๋ธ” ko-trocr (์ƒ๋Œ€ ๊ฐ€์šด๋ฐ 1๋“ฑ)
๋‚˜๋ˆ”ํŽœ์Šคํฌ๋ฆฝํŠธ (NanumPenScript) 112 95.4 95.4 96.2 69.1
๊ฐ์ž๊ฝƒ (GamjaFlower) 112 97.2 97.9 97.9 82.1
ํ•˜์ด๋ฉœ๋กœ๋”” (HiMelody) 112 97.3 97.2 97.3 75.3
์‹ฑ๊ธ€๋ฐ์ด (SingleDay) 112 98.2 99.0 98.9 88.9
๋™ํ•ด๋…๋„ (EastSeaDokdo) 97 95.0 93.1 94.1 65.9
๊ธฐ๋ž‘ํ•ด๋ž‘ (KirangHaerang) 95 99.0 97.9 97.8 87.9

์‚ฌ์ง„๋งˆ๋‹ค (์ค„ ์นธ, ์ž๋ชจ ๋‹ฎ์Œ)

01 02 03 04 05 06 07 08 09 10 11
์ด ๋ชจ๋ธ 94 100 98 82 100 94 89 100 88 91 93
41M ๋‹จ์ผ ๋ชจ๋ธ 94 100 100 96 93 83 94 100 89 87 100
41M ์•™์ƒ๋ธ” 100 100 100 98 100 90 95 97 91 93 98

์‚ฌ์ง„์€ ์žฅ๋งˆ๋‹ค ์„œ๋„ˆ ์ค„๋ฟ์ด๋ผ ํ•œ ์ค„์ด ํ•œ ์žฅ ์ ์ˆ˜๋ฅผ 10์  ๋„˜๊ฒŒ ์›€์ง์ž…๋‹ˆ๋‹ค. ๊ทธ๋ž˜์„œ ์‚ฌ์ง„ ์ชฝ ์ฐจ์ด๋Š” ๊ธ€์”จ์ฒด ์ชฝ๋ณด๋‹ค ํ›จ์”ฌ ๋œ ๋ฏฟ์„ ๋งŒํ•ฉ๋‹ˆ๋‹ค.

๋น ๋ฅด๊ธฐ๋Š” ์ด๋ ‡๊ฒŒ ์Ÿ€์Šต๋‹ˆ๋‹ค

  • ์„œ๋ฅธ ์ค„์งœ๋ฆฌ ์„œ๋ฅ˜ ํ•œ ์ชฝ์„ ๊ธฐ์ค€์œผ๋กœ ์žฝ๋‹ˆ๋‹ค. ์ค„ ์ž๋ฅด๊ธฐ(CPU, ๋ชจ๋‘ ๊ฐ™์€ 0.40์ดˆ) + ์ฝ๊ธฐ์ž…๋‹ˆ๋‹ค.
  • GPU ๋Š” ๋ฌถ์Œ์„ 1ยท4ยท8ยท16ยท32์ค„๋กœ ๋ฐ”๊ฟ” ๊ฐ€๋ฉฐ ์„ธ ๋ฒˆ์”ฉ ์žฌ๊ณ , ์นด๋“œ(8GB) ๋ฐ–์œผ๋กœ ๋„˜์น˜์ง€ ์•Š์€ ๊ฒƒ ๊ฐ€์šด๋ฐ ์ค„๋‹น ๊ฐ€์žฅ ์‹ผ ๋ฌถ์Œ์„ ์”๋‹ˆ๋‹ค. ์ด ๋ชจ๋ธ์€ 16์ค„ ๋ฌถ์Œ์ด ๊ฐ€์žฅ ์‹ธ๊ณ , ๊ทธ๋•Œ VRAM ๊ผญ๋Œ€๊ธฐ๋Š” 2.2GB ์ž…๋‹ˆ๋‹ค.
  • CPU ๋Š” 1ยท8์ค„ ๋ฌถ์Œ, ํฐ VL ๋ชจ๋ธ์€ ์ค„๋‹น 1์ดˆ ์•ˆํŒŽ์ด๋ผ 1์ค„๋งŒ ์žฝ๋‹ˆ๋‹ค.
  • CPU ๊ฐ’์€ ์ด ๋…ธํŠธ๋ถ์˜ ๊ทธ๋•Œ ์ƒํƒœ๋ฅผ ํฌ๊ฒŒ ํƒ‘๋‹ˆ๋‹ค. ์ค„/์ดˆ ์ž์ฒด๋ณด๋‹ค ๊ฐ™์€ ํŒ ์•ˆ์˜ ์ˆœ์„œ๋ฅผ ๋ด ์ฃผ์„ธ์š”.

๊ฐ™์ง€ ์•Š์€ ๊ฒƒ โ€” ๋ˆ„๊ตฌ์—๊ฒŒ ์œ ๋ฆฌํ•œ์ง€๊นŒ์ง€

๋ฌด์—‡์ด ์–ด๋–ป๊ฒŒ ๋ˆ„๊ตฌ์—๊ฒŒ ์œ ๋ฆฌ
์‹œํ—˜ ๊ธ€๊ผด ์šฐ๋ฆฌ ํ•™์Šต์—์„œ๋งŒ ๋บ์Šต๋‹ˆ๋‹ค. ์ƒ๋Œ€๊ฐ€ ๋ฌด์—‡์„ ๋ณด๊ณ  ๋ฐฐ์› ๋Š”์ง€๋Š” ๋ชจ๋ฆ…๋‹ˆ๋‹ค ์ƒ๋Œ€
์ค„ ์ž๋ฅด๊ธฐ ์ค„ ์ธ์‹๊ธฐ๋“ค์—๊ฒŒ ์šฐ๋ฆฌ ์ค„ ์ž๋ฅด๊ธฐ๋ฅผ ๋นŒ๋ ค์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค ์ƒ๋Œ€
๋ฌผ์Œ Qwen3-VL-2B ๋Š” ์ •ํ•ด ๋‘” OCR ๋ฌผ์Œ์ด ์—†์–ด ์…‹์„ ๋Œ€ ๋ณด๊ณ  ๊ฐ€์žฅ ์ž˜ ์ฝ๋Š” ๊ฒƒ์„ ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค ์ƒ๋Œ€
์‚ฌ์ง„ ํ†ต์งธ ๊ทธ๋ฆผ ํฌ๊ธฐ ์šฐ๋ฆฌ์™€ ์›๋ณธ์€ 4,096ํ† ํฐ๊นŒ์ง€๋กœ ์ค„์ด๊ณ , ์ƒ๋Œ€ VL ๋‘˜์€ ๊ทธ์ชฝ ๊ธฐ๋ณธ ํฌ๊ธฐ ๊ทธ๋Œ€๋กœ ์ƒ๋Œ€
๊ธธ์ด ์ƒํ•œ ko-trocr ์˜ ๊ธฐ๋ณธ 16๊ธ€์ž๋ฅผ 64๋กœ ํ’€์–ด ์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค ์ƒ๋Œ€
๊ทธ๋ฆฌ๊ฐœ ์šฐ๋ฆฌ๋Š” ์‹œํ—˜์ง€์™€ ๊ฐ™์€ ๊ทธ๋ฆฌ๊ฐœ๋กœ ๋งŒ๋“  ํ•ฉ์„ฑ ๊ทธ๋ฆผ์œผ๋กœ ๋ฐฐ์› ์Šต๋‹ˆ๋‹ค(๊ธ€๊ผด์€ ๋‹ค๋ฆ„). ์‹œํ—˜ ๊ธ€์”จ์ฒด ์ชฝ์€ ์šฐ๋ฆฌ์—๊ฒŒ ๋‚ฏ์ต์€ ๊ผด์ž…๋‹ˆ๋‹ค โ€” ๊ทธ๋ž˜์„œ ์‚ฌ์ง„ ์ชฝ์ด ๊ฐ€์žฅ ๊นจ๋—ํ•œ ๊ฒฌ์คŒ์ž…๋‹ˆ๋‹ค ์šฐ๋ฆฌ
๋งŒ๋“  ๋ชฉ์  ์ƒ๋Œ€ VL ์€ ๋ฌธ์„œ ์ „์ฒดยท์—ฌ๋Ÿฌ ์–ธ์–ด์šฉ์ž…๋‹ˆ๋‹ค. ์ด๊ฒƒ์€ ํ•œ๊ธ€ ์†๊ธ€์”จ ํ•˜๋‚˜๋งŒ ๋…ธ๋ ธ์Šต๋‹ˆ๋‹ค ์šฐ๋ฆฌ

์ƒ๋Œ€๋“ค์ด ๋‚˜์œ ๋ชจ๋ธ์ด๋ผ๋Š” ๋œป์ด ์•„๋‹™๋‹ˆ๋‹ค. ์ธ์‡„ ๋ฌธ์„œยทํ‘œยท์žฅ๋ฉด ๊ธ€์ž๋ฅผ ์ฝ์œผ๋ผ๊ณ  ๋งŒ๋“  ๋ชจ๋ธ์— ์†๊ธ€์”จ๋ฅผ ์ค€ ๊ฒƒ์ด๊ณ , ๊ทธ ๋ฐ˜๋Œ€(์ธ์‡„ ๋ฌธ์„œ ์ „์ฒด, ํ‘œ)๋Š” ์žฌ์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค. ์žฌ๋ฉด ์ด ๋ชจ๋ธ์ด ์งˆ ๊ฐ€๋Šฅ์„ฑ์ด ํฝ๋‹ˆ๋‹ค.

LM StudioยทOllamaยทllama.cpp ์ค„์€ ๊ฐ™์€ Q8_0 ํŒŒ์ผ์„ ๋ฐ›์€ ์„ค์ • ๊ทธ๋Œ€๋กœ ๋„์šฐ๊ณ , ๊ฐ™์€ ์‹œํ—˜์ง€(๊ธ€์”จ์ฒด ๋ฒŒ๋งˆ๋‹ค 100์ค„)์™€ ๊ฐ™์€ ์‚ฌ์ง„์œผ๋กœ OpenAI ๊ผด API ์— ๋ฌผ์–ด ์žฐ ๊ฐ’์ž…๋‹ˆ๋‹ค. LM StudioยทOllama ๊ฐ€ ์ด ์ €์žฅ์†Œ์—์„œ ๋ฐ›์€ Q8_0 ๊ณผ ๊ทธ๋ฆผ ํŒŒ์ผ(mmproj)์€ ์˜ฌ๋ฆฐ ํŒŒ์ผ๊ณผ sha256 ์ด ๊ฐ™์Šต๋‹ˆ๋‹ค.

์“ฐ๋Š” ๋ฒ•

๋ฌผ์Œ์€ ๋ฌด์—‡์ด๋“  ๋ฉ๋‹ˆ๋‹ค(ใ€ŒOCRใ€, ใ€Œ๊ธ€์ž ์ฝ์–ด ์ค˜ใ€, ์˜์–ด, ๋นˆ ๋ฌผ์Œ). ๋Œ€๋‹ต์€ ๊ธ€์ž๋งŒ ์ค„๋งˆ๋‹ค ํ•œ ์ค„๋กœ ๋‚˜์˜ต๋‹ˆ๋‹ค. ์ƒ๊ฐ ์นธ์€ ๋Œ€ํ™” ํ‹€์—์„œ ๋Š˜ ๋‹ซํ˜€ ์žˆ์–ด์„œ ์ƒ๊ฐํ•˜๋Š” ๋ฐ ์‹œ๊ฐ„์„ ์“ฐ์ง€ ์•Š์Šต๋‹ˆ๋‹ค. temperature ๋Š” 0 ์œผ๋กœ ๋‘์‹ญ์‹œ์˜ค(OCR ์€ ๊ณ ๋ฅด๊ฒŒ ๋ฝ‘์œผ๋ฉด ํ‹€๋ฆฝ๋‹ˆ๋‹ค).

Ollama

ollama run hf.co/localdeel/ko-hand-ocr-vl-GGUF:Q8_0

๊ทธ๋ฆผ ์ฝ๋Š” ๋ชซ๊ณผ ์„ค์ •(temperature 0, top_k 1, ๋งฅ๋ฝ 8192)์ด ๊ฐ™์ด ๋ฐ›์•„์ง‘๋‹ˆ๋‹ค(์ €์žฅ์†Œ์˜ params).

API ๋กœ๋Š” ๊ทธ๋ฆผ์„ base64 ๋กœ ๋„ฃ์Šต๋‹ˆ๋‹ค.

curl http://localhost:11434/api/chat -d '{"model": "hf.co/localdeel/ko-hand-ocr-vl-GGUF:Q8_0", "stream": false,
  "messages": [{"role": "user", "content": "OCR", "images": ["<base64>"]}]}'

LM Studio

๊ฒ€์ƒ‰์—์„œ ko-hand-ocr-vl ์„ ์ฐพ์•„ Q8_0 ์„ ๋ฐ›๊ฑฐ๋‚˜, ๋ช…๋ น์ค„๋กœ ๋ฐ›์Šต๋‹ˆ๋‹ค.

lms get https://huggingface.co/localdeel/ko-hand-ocr-vl-GGUF@q8_0

๊ทธ๋ฆผ ์ฝ๋Š” ๋ชซ(mmproj)์ด ๊ฐ™์ด ๋ฐ›์•„์ง€๊ณ (ํ•ฉ์ณ 765MB), ์„ค์ •์„ ํ•˜๋‚˜๋„ ์•ˆ ๋ฐ”๊ฟ”๋„ ๋งž๊ฒŒ ๋•๋‹ˆ๋‹ค โ€” ์ƒ๊ฐ ์นธ์€ ๋Œ€ํ™” ํ‹€์—์„œ ๋‹ซํ˜€ ์žˆ๊ณ , ๊ฐ€์žฅ ๊ทธ๋Ÿด๋“ฏํ•œ ๊ธ€์ž ํ•˜๋‚˜๋งŒ ๊ณ ๋ฅด๋Š” ์„ค์ •(top_k 1)์ด ํŒŒ์ผ์— ๋“ค์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ์œ„ ํ‘œ์˜ LM Studio ์ค„์ด ๊ทธ๋ ‡๊ฒŒ ๋ฐ›์€ ๊ทธ๋Œ€๋กœ ์žฐ ๊ฐ’์ž…๋‹ˆ๋‹ค. ๋Œ€ํ™”์ฐฝ์— ์‚ฌ์ง„์„ ๋Œ์–ด ๋†“๊ฑฐ๋‚˜, ์„œ๋ฒ„๋ฅผ ์ผœ๊ณ  OpenAI ๊ผด๋กœ ๋ถ€๋ฆ…๋‹ˆ๋‹ค.

llama.cpp

llama-server -hf localdeel/ko-hand-ocr-vl-GGUF:Q8_0 --temp 0

vLLM

vllm serve localdeel/ko-hand-ocr-vl --max-model-len 8192 --limit-mm-per-prompt '{"image": 1}'

๊ตฌ์กฐ๊ฐ€ Qwen3.5-0.8B ๊ทธ๋Œ€๋กœ๋ผ vLLM ์˜ Qwen3.5 ์ง€์›์œผ๋กœ ์˜ฌ๋ผ๊ฐ‘๋‹ˆ๋‹ค. ์ด ๋ชจ๋ธ์„ ๋งŒ๋“  PC ์—์„œ๋Š” vLLM ์ด ๋Œ์ง€ ์•Š์•„(Windows) ์ง์ ‘ ๋„์›Œ ๋ณด์ง€๋Š” ๋ชปํ–ˆ์Šต๋‹ˆ๋‹ค.

transformers

from transformers import AutoModelForImageTextToText, AutoProcessor
from PIL import Image

proc = AutoProcessor.from_pretrained("localdeel/ko-hand-ocr-vl")
model = AutoModelForImageTextToText.from_pretrained("localdeel/ko-hand-ocr-vl", dtype="auto", device_map="auto")
talk = [{"role": "user", "content": [{"type": "image"}, {"type": "text", "text": "OCR"}]}]
text = proc.apply_chat_template(talk, add_generation_prompt=True)
ask = proc(text=[text], images=[Image.open("photo.jpg")], return_tensors="pt").to(model.device)
out = model.generate(**ask, max_new_tokens=512)
print(proc.batch_decode(out[:, ask["input_ids"].shape[1]:], skip_special_tokens=True)[0])

OpenAI ๊ผด ์„œ๋ฒ„(LM Studioยทllama-serverยทOllamaยทvLLM)์—๋Š” ๋ชจ๋‘ ๊ฐ™์€ ๊ผด๋กœ ๋ฌป์Šต๋‹ˆ๋‹ค.

import base64
from openai import OpenAI

client = OpenAI(base_url="http://localhost:1234/v1", api_key="none")   # ์„œ๋ฒ„ ์ฃผ์†Œ๋งŒ ๋ฐ”๊พผ๋‹ค
data = base64.b64encode(open("photo.jpg", "rb").read()).decode()
got = client.chat.completions.create(model="ko-hand-ocr-vl", temperature=0, messages=[{
    "role": "user", "content": [
        {"type": "image_url", "image_url": {"url": "data:image/jpeg;base64," + data}},
        {"type": "text", "text": "OCR"}]}])
print(got.choices[0].message.content)

์–ด๋–ป๊ฒŒ ๋งŒ๋“ค์—ˆ๋‚˜

  • ์›๋ณธ: Qwen3.5-0.8B (Gated DeltaNet 18์ธต + ์ฃผ์˜ 6์ธต, ๊ทธ๋ฆผ ์ธ์ฝ”๋” 1์–ต ๊ฐœ)
  • LoRA(r=64, ๋ชจ๋“  ์„ ํ˜• ์ธต โ€” ๊ทธ๋ฆผ ์ธ์ฝ”๋” ํฌํ•จ), ๋Œ€๋‹ต ์ž๋ฆฌ์—์„œ๋งŒ ์†์‹ค, ํ•™์Šต๋ฅ  5e-5. ๋ˆ„์  46,000๊ฑธ์Œ(ํ•œ ๊ฑธ์Œ = ์ค„ 6๊ฐœ ๋˜๋Š” ์‚ฌ์ง„ 1์žฅ)์„ ์—ฌ๋Ÿฌ ์ฐจ๋ก€ ์ด์–ด ๋Œ๋ ธ๊ณ , ๊ธด ๋‘ ์ฐจ๋ก€(20,000๊ฑธ์Œ์”ฉ) ๋๋งˆ๋‹ค ๋งˆ์ง€๋ง‰ ์ €์žฅ ๋„ท์„ ํ‰๊ท ํ–ˆ์Šต๋‹ˆ๋‹ค. ๋…น์—ฌ์„œ ๋‚ด๋ณด๋ƒ„
  • ํ•™์Šต ๊ทธ๋ฆผ์€ ์ „๋ถ€ ํ•ฉ์„ฑ: ์†๊ธ€์”จ ๊ธ€๊ผด 450๋ฒŒ๋กœ ๊ทธ๋ฆฐ ํ•œ ์ค„(55%), ํ•œ ์‚ฌ๋žŒ์ด ์ข…์ด์— ์“ด 110์ค„์„ ์ฑ…์ƒ ์œ„์—์„œ ์ฐ์€ ๊ฒƒ์ฒ˜๋Ÿผ ๋งŒ๋“  ์‚ฌ์ง„(35%, ๊ธด ๋ณ€ 6402,400px), ์ธ์‡„์ฒด ์ค„(10%)
  • ์ง€์šด ์ž๊ตญยทํš ํœ˜๊ธฐยท๊ธฐ์šธ์ž„ยท๊ทธ๋Š˜ยทํ๋ฆผยทJPEG ๋ฅผ ์„ž์—ˆ์Šต๋‹ˆ๋‹ค. ์‹œํ—˜ ๊ธ€์”จ์ฒด 30๋ฒŒ๊ณผ ์‚ฌ์ง„์€ ํ•™์Šต์—์„œ ๋บ์Šต๋‹ˆ๋‹ค
  • ์‚ฌ์ง„์ฒ˜๋Ÿผ ๋งŒ๋“  ์ชฝ์€ ํ•œ ๊ฐ€๋กœ์ค„์— ๋‘์„ธ ๋ฉ์ด๋ฅผ ๋‚˜๋ž€ํžˆ ์“ฐ๊ธฐ๋„ ํ•ฉ๋‹ˆ๋‹ค(ใ€ŒํŒ€: โ€ฆใ€ ์™ผ์ชฝ, ใ€Œ์ด๋ฆ„: โ€ฆใ€ ์˜ค๋ฅธ์ชฝ). ์‹ค์ œ ์„œ์‹์ด ๊ทธ๋ ‡๊ฒŒ ์“ฐ์—ฌ์„œ, ์•ˆ ๊ทธ๋Ÿฌ๋ฉด ์˜ค๋ฅธ์ชฝ ๋ฉ์ด๋ฅผ ๋น ๋œจ๋ ธ์Šต๋‹ˆ๋‹ค
  • ๊ธ€๊ผด์ด ๋ชป ๊ทธ๋ฆฌ๋Š” ๊ธ€์ž(โ–กยท๋นˆ์นธ)๊ฐ€ ๋“  ์ค„์€ ๊ทธ ๊ธ€๊ผด๋กœ ๊ทธ๋ฆฌ์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ๋„ค๋ชจ๋ฅผ ๊ทธ๋ ค ๋†“๊ณ  ์ •๋‹ต์— ์ œ ๊ธ€์ž๋ฅผ ์ ์œผ๋ฉด ๋ชป ์ฝ๋Š” ์ž๋ฆฌ๋ฅผ ์ง€์–ด๋‚ด๋ผ๊ณ  ๊ฐ€๋ฅด์น˜๋Š” ๊ผด์ด๋ผ์„œ์ž…๋‹ˆ๋‹ค
  • ๋‘˜์งธ ์ฐจ๋ก€(20,000๊ฑธ์Œ)์—๋Š” ๋œป ์—†๋Š” ๊ธ€์ž ์ค„์„ 15% ์„ž์—ˆ์Šต๋‹ˆ๋‹ค. ํ๋ฆฐ ๊ธ€์ž๋ฅผ ํ”ํ•œ ๋‚ฑ๋ง๋กœ ๊ณ ์ณ ์ฝ๋Š” ๋ฒ„๋ฆ‡(์ธ์žฌํŒ€โ†’์ธ์‚ฌํŒ€)์„ ์ค„์ด๋ ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค
  • ๋ฐ”๊พผ ์„ค์ •: ๋Œ€ํ™” ํ‹€์˜ ์ƒ๊ฐ ์นธ์„ ๋Š˜ ๋‹ซ์Œ, ๊ทธ๋ฆผ ์ƒํ•œ 4,096ํ† ํฐ, top_k: 1

ํ•œ๊ณ„

  • ์‚ฌ๋žŒ ์†๊ธ€์”จ๋กœ๋Š” ํ•™์Šตํ•˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค. ๊ธ€๊ผด๋กœ ๊ทธ๋ฆฐ ์†๊ธ€์”จ๋งŒ ๋ดค์Šต๋‹ˆ๋‹ค. ํ˜๋ ค ์“ด ๊ธ€์”จ, ๊ฒน์ณ ์“ด ๊ธ€์”จ๋Š” ์•ฝํ•ฉ๋‹ˆ๋‹ค.
  • ์‹œํ—˜ ์‚ฌ์ง„์ด 11์žฅ๋ฟ์ด๋ผ ์‚ฌ์ง„ ์ ์ˆ˜์˜ ํญ์ด ํฝ๋‹ˆ๋‹ค(ํ•œ ์žฅ์ด 9%p).
  • ํ‘œยท๋„์žฅยท์„œ๋ช… ๊ฐ™์€ ์„œ์‹ ๊ตฌ์กฐ๋Š” ์ฝ์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ๊ธ€์ž๋งŒ ์ค„๋งˆ๋‹ค ๋ƒ…๋‹ˆ๋‹ค.
  • safetensors ์˜ mtp.*(์—ฌ๋Ÿฌ ํ† ํฐ ๋ฏธ๋ฆฌ ์ง์ž‘) ๋จธ๋ฆฌ๋Š” ์›๋ณธ ๊ทธ๋Œ€๋กœ๋ผ ๋ฏธ์„ธ์กฐ์ •๋˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค. ๋‹ต์€ ์•ˆ ๋ฐ”๋€Œ์ง€๋งŒ ์ง์ž‘์ด ๋œ ๋งž์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. GGUF ์—๋Š” ์—†์Šต๋‹ˆ๋‹ค.
  • safetensors ๋ฅผ ์†์ˆ˜ GGUF ๋กœ ๋ฐ”๊ฟ€ ๋•Œ llama.cpp convert_hf_to_gguf.py ๊ฐ€ ํ† ํฌ๋‚˜์ด์ €๋ฅผ ๋ชป ์•Œ์•„๋ด…๋‹ˆ๋‹ค โ€” ์–ดํœ˜๋ฅผ ์ž˜๋ผ ์ง€๋ฌธ์ด ๋‹ฌ๋ผ์กŒ๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. ์›๋ณธ๊ณผ ๊ฐ™์€ qwen35 ๋กœ ์ •ํ•ด ์ฃผ๊ณ  ๋ฐ”๊พธ๊ฑฐ๋‚˜ GGUF ํŒ์„ ๋ฐ›์Šต๋‹ˆ๋‹ค.

๋ผ์ด์„ ์Šค

Apache-2.0. ์›๋ณธ Qwen3.5-0.8B ๋„ Apache-2.0 ์ž…๋‹ˆ๋‹ค. ์ž์„ธํ•œ ๊ฒƒ์€ NOTICE.

English

An open-weight Korean handwriting OCR model: give it a photo, get the text back. It is Qwen3.5-0.8B fine-tuned on Korean handwriting with the architecture unchanged, so anything that runs Qwen3.5 โ€” Ollama, LM Studio, llama.cpp, vLLM, transformers โ€” runs it with no new code.

2026-10-12 โ€” vocabulary cut, same answers. Of the base Qwen3.5 vocabulary (248,077 tokens) we kept Hangul, jamo, digits, symbols, short Latin pieces and every token our training and test text uses: 59,453 tokens (853M -> 660M parameters). The weights were not retrained, and on the test fonts, the held-out fonts, the photo line crops and the whole photos every answer came out identical, character for character โ€” through transformers and through llama-server (Q8_0, one request at a time). Downloads shrink: safetensors 1.63GB -> 1.27GB, Q8_0 774MB -> 570MB, Q4_K_M 505MB -> 345MB, and llama.cpp generates 26โ€“36% faster (this PC, llama-bench). On 200 lines built only from long English words that left the vocabulary, 173 came out the same and jamo similarity went 98.34 -> 98.01. The line table and the speeds were re-measured on the cut model; the whole-photo and runtime tables were measured before the cut (the answers are the same, so they stand). The uncut model is at this repo's tag full-vocab.

All numbers below are measured on the same test sheets on the same PC. Test fonts were held out of training; the 11 photos are real handwriting, never used for training, and not published.

Character similarity, ten models on the same test sheets

Model Size Test fonts 6 (char. sim. / CER) Unseen fonts 24 Photos 11, line crops (sim. / CER) GPU lines/s
ko-hand-ocr-vl (this, 660M) 1.27GB 97.0 / 5.8 97.4 93.4 / 15.8 10.8
Qwen3.5-0.8B (base, before fine-tuning) 1.63GB 53.0 / 161.8 51.4 54.2 / 115.3 9.6
ko-hand-ocr single model (41M, lines only) 78MB 96.8 / 7.2 97.3 94.2 / 13.4 42.0
ko-hand-ocr ensemble (3 models) 225MB 97.1 / 6.7 97.7 96.5 / 8.6 8.3
Qwen/Qwen3-VL-2B-Instruct 3.96GB 66.6 / 79.1 68.0 69.9 / 38.3 7.3
PaddlePaddle/PaddleOCR-VL-1.6 1.79GB 63.6 / 62.9 66.5 61.7 / 44.0 1.2
ddobokki/ko-trocr 408MB 78.2 / 42.3 80.4 78.5 / 39.2 6.1
PaddleOCR PP-OCRv5 (korean) 13MB 72.7 / 40.5 75.3 71.8 / 39.7 397.4
EasyOCR (korean) 15MB 56.6 / 64.3 64.6 66.9 / 55.0 81.3
Tesseract 5 (kor+eng) 6MB 17.7 / 97.3 19.7 13.2 / 98.6 โ€”

Whole photos (original size, capped at 4,096 image tokens). Page compares the whole page's text joined together (the usual page-level CER); line-paired matches each line read to a label line. The two rival VLMs get their own default image size, not our cap:

Whole photos

Model Page char. similarity Page CER Worst photo (page) Line-paired similarity
ko-hand-ocr-vl (this) 94.3 15.3 78.3 95.6
Qwen3.5-0.8B (base) 72.7 42.1 55.3 62.1
Qwen3-VL-2B-Instruct 76.9 41.1 47.6 72.5
PaddleOCR-VL-1.6 79.9 37.3 54.5 57.3

Same model, different runtimes (100 lines per font). The llama.cpp family is lower on photos because of image preprocessing, not quantization (the unquantized F16 GGUF scores the same): llama.cpp keeps the aspect ratio and pads the edge with black, and shrinks small images down to 8 tokens, while this model was trained from 64. The mmproj now carries the 64-4,096 token limits (2026-10-09), which llama.cpp and LM Studio read on load: photo line crops went 90.8 -> 92.5 in the same LM Studio and 90.9 -> 92.0 in llama-server. Ollama 0.34.4 does not read them yet. The black padding cannot be changed from the file.

How it runs Test fonts 6 Photos, line crops Photos, whole
transformers (bf16, safetensors) 96.9 93.3 94.3
llama.cpp llama-server โ€” Q8_0 96.7 92.0 91.8
LM Studio โ€” Q8_0, ๋ฐ›์€ ๊ทธ๋Œ€๋กœ (์„ค์ • ์•ˆ ๊ฑด๋“œ๋ฆผ) 96.4 92.5 91.8
Ollama โ€” hf.co/โ€ฆ:Q8_0, ๋ฐ›์€ ๊ทธ๋Œ€๋กœ 96.2 90.9 91.8

How the benchmark was run

  • Same test sheets, same PC, all ten models in one run (2026-10-13 14:59). GPU: NVIDIA GeForce RTX 5060 Laptop GPU (8GB); CPU: AMD Ryzen 7 260 (laptop, 8 cores); Windows 11; PyTorch 2.14.0+cu130.
  • Test fonts (6): 640 lines / 4,448 characters drawn with six free handwriting fonts (OFL) held out of training, with paper texture, tilt and strike-through marks; fixed seed never used in training. About a third of the lines are meaningless syllable strings (no language prior to lean on), a third are form fields, the rest titles and sentences, with digits, dates, Latin abbreviations and symbols mixed in. Lines a font cannot draw are dropped.
  • Unseen fonts (24): 634 lines, never used for training or for picking models.
  • Photos (11): real handwriting on real forms, 34 lines / 209 characters, never used for training and not published. Line crops come from the ko-hand-ocr line cutter, so every model gets the same crops. A 12th photo (low-res printed form + signatures) is measured but not scored.
  • Scores: character similarity = difflib ratio over jamo with spaces removed, averaged per line; CER = edit distance / reference characters, pooled over all characters (can exceed 100% when a model rambles).
  • Rivals run exactly as their model cards say, in their own environments; timing counts only time inside the engine. Qwen3-VL-2B got the best of three prompts; ko-trocr's 16-char length cap was raised to 64; the rival VLMs read whole photos at their own default size. Our training images come from the same renderer as the font sheets (different fonts), so the photo set is the cleanest comparison.
  • Runtimes: llama.cpp build 11474, LM Studio (llama.cpp CUDA 12 engine 2.55.0) and Ollama 0.34.4, each with the downloaded Q8_0 file and default settings, 100 lines per font. The Q8_0 and mmproj files LM Studio and Ollama downloaded from this repo match the upload by sha256.

Test sheet

Usage

Any prompt works ("OCR", "Read the text", Korean, or empty). The answer is only the text, one line per line. The thinking block is always closed by the chat template, so no time is spent thinking. Use temperature 0 (sampling makes OCR worse).

  • Ollama: ollama run hf.co/localdeel/ko-hand-ocr-vl-GGUF:Q8_0
  • LM Studio: search ko-hand-ocr-vl or lms get https://huggingface.co/localdeel/ko-hand-ocr-vl-GGUF@q8_0. The mmproj vision file comes with it and it works with no settings changed (thinking is closed in the template; top_k 1 is in the file).
  • llama.cpp: llama-server -hf localdeel/ko-hand-ocr-vl-GGUF:Q8_0 --temp 0
  • vLLM: vllm serve localdeel/ko-hand-ocr-vl --max-model-len 8192 โ€” same architecture as Qwen3.5-0.8B, so vLLM's Qwen3.5 support loads it. Not run by us (the build machine is Windows).
  • transformers: AutoModelForImageTextToText + AutoProcessor, as in the Korean section.

Limits. Trained on synthetic handwriting only (450 fonts), no real handwriting samples. Only 11 test photos, so the photo scores are noisy. Reads text line by line; does not parse tables or forms. To convert the safetensors to GGUF yourself, set the pre-tokenizer to qwen35 (the cut vocabulary changes the fingerprint convert_hf_to_gguf.py checks); the GGUF repo already has it. License: Apache-2.0 (same as the base model).

Downloads last month
130
Safetensors
Model size
0.7B params
Tensor type
F32
ยท
BF16
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for localdeel/ko-hand-ocr-vl

Finetuned
(843)
this model
Quantizations
1 model