Models by script

Character error rate for a selection of the handwriting systems we tested, on 8,136 lines across seven script types, plus an estimate for Najdi. Lower is better.

Kindi v0.5 preview has the lowest error on six of the seven scripts. On Modern handwriting, other systems we tested read better, including sherif-v3.

System	Oriental1,528	Ruq‘ah2,096	Nastaliq663	Ta‘liq182	Naskh1,205	Maghrebi1,071	Modern1,391	All8,136	Najdiestimate	Lines read
Read every line (8,136)
Kindi v0.5 preview	3.70	5.82	8.18	8.83	9.82	10.43	15.30	8.96†	12.4	8,136
HATFormer	32.03	17.91	20.73	13.43	21.19	33.90	41.42	26.81	30.0	8,136
sdkv2	24.74	21.93	32.46	35.09	40.03	48.78	39.65	33.72	40.4	8,136
Chandra OCR 2	29.33	34.33	48.14	27.50	33.92	52.07	33.88	37.07	30.5	8,136
PaddleOCR-VL 1.6	35.95	44.13	36.06	27.37	49.31	52.41	36.87	42.90	47.6	8,136
Qwen3.5-4B	36.83	48.90	44.73	34.09	48.82	55.94	36.12	45.45	41.9	8,136
olmOCR 2	36.90	48.43	54.16	34.43	45.06	58.70	36.64	45.91	45.3	8,136
dots.mocr	37.86	49.64	37.69	27.33	41.65	57.98	50.69	46.68	38.5	8,136
Qari-OCR v0.4	40.00	50.32	43.66	33.60	49.98	56.91	40.79	47.30	47.8	8,136
EasyOCR	48.80	52.19	42.70	35.30	50.81	55.42	45.49	49.59	54.7	8,136
LightOnOCR 2	43.47	53.89	70.73	48.49	54.11	62.49	39.40	52.55	49.4	8,136
Qari-OCR v0.3	40.02	58.08	67.00	33.13	58.19	57.57	45.26	53.78	43.2	8,136
Tesseract	68.72	63.70	70.32	49.22	68.76	65.11	51.58	63.38	54.8	8,136
DeepSeek-OCR 2	73.88	69.82	70.65	59.91	70.21	74.46	60.19	69.07	66.4	8,136
GLM-OCR	83.71	79.55	81.95	78.28	83.19	83.97	81.39	81.82	67.2	8,136
Read part of the test
sherif-v3	15.94	18.42*	31.02	19.39	34.98*	34.41*	7.88‡	—	27.9	7,838
Gemma 4 31B	21.87	26.55*	17.16	14.49	24.69*	42.13*	24.10	—	28.4	7,838
Baseer Nakba	19.57	23.34*	33.31	27.68	41.26*	53.67*	28.19	—	32.7	7,838
deep.sa OCR v2.0	—	49.25*	44.07	37.03	51.23*	52.57*	—	—	—	4,458
Mistral OCR v4.1	—	29.33*	40.21*	29.37	39.82*	56.65*	—	—	—	2,000
Qwen2.5-VL 3B	—	43.68*	58.45*	37.49	49.58*	59.04*	—	—	41.5	2,000

How to read the table

Bold numbers: The lowest error in the column among the systems listed that read all of its lines. Modern has none, because systems not listed here read it better.
* Partial read: The system read only some of that script's lines, so compare its number with care.
—: No score: the system hasn't read these lines.
‡ Trained on these lines: sherif-v3 trained on most of the Modern lines, so its score there is better than it would be on new handwriting.
† Calculated: Kindi v0.5 preview's All score is worked out from its seven script scores, weighted by characters.
Najdi: An estimate, not a test result. These pages have no line-by-line reference, so each read is compared with its closest stretch of a printed transcription. Every system is scored on the same 76 lines, and lines with no close match are dropped, which favours weak systems. DeepSeek-OCR 2, GLM-OCR and Tesseract matched very few lines, so their numbers say little.
Qwen3.5-4B: It read the seven scripts with our prompt and Najdi with a plainer one.
Why it isn't 8.83%: The 8.83% on the Kindi HTR page comes from a narrower test: 2,000 lines in five scripts. This table uses the wider test.

Back to Kindi HTR