Persian OCR Triple Threat Benchmark

Sortable leaderboard — ranked by Overall Triple Threat

1	Chandra OCR 2 NVFP4A16	35.6M	104.92	10.21	12.12	217.47	166.86	11.19	185.39	165.36	24.05	vLLM 0.20.1 · BF16 · greedy	Bina · Surya	complete	Checkpoint revision aaa02f3811945a91062062994c5c4a3f4c0af2b0; official Free OCR. prompt; 8× RTX PRO 6000 Blackwell 96 GB, one replica per GPU; batch size 64 per GPU; max new tokens 2048; n-gram repetition processor size 40. Dataset revision fefe25215114da8ac0ea21ff5e8f97204b2762d5 was streamed by HTTP Parquet range reads without materializing it. All 13,338 rows scored with zero missing, blank, or failed predictions; 1,844 outputs reached the token cap and were included as generated.


1	🥇 Bina 0.1 Koochik	666M	9.84	10.21	12.12	10.83	6.54	11.19	11.41	8.38	9.22		Bina · Surya	complete	Persian LoRA checkpoint step 8000, merged for inference; batch size 8 per GPU; max new tokens 4096. All 13,338 rows scored with zero missing predictions. Triple Threat uses the leaderboard's current 20% WER · 20% CER · 60% S³ weighting.
2	🥈 Bina 0.2 Rizeh	35.6M	11.47	11.78	13.42	8.93	2.35	9.85	19.08	13.64	13.7	Paddle Inference · CTC	Bina	complete	PP-OCRv6_medium_rec Stage-8 release with the adaptive75 two-layer full-page pipeline. All 13,338 rows scored with zero missing predictions.
3	🥉 🔒 Gemini 3.5 Flash		15.72	13.97	15.74	6.71	3.68	3.89	33.95	29.06	24.05		Gemini	complete	OpenRouter routing; minimal reasoning effort; reasoning excluded from returned transcription.
4	🔒 Gemini 3.1 Flash Lite		18.62	16.92	18.79	7.22	4.43	6.59	39.5	33.54	27.24		Gemini	complete	OpenRouter routing; default reasoning effort; reasoning excluded from returned transcription.
5	Bina 0.2 RizehPizeh	5.29M	25.17	27.41	30.31	25.22	8.81	27.22	33.13	20.15	27.59	Paddle Inference · CTC	Bina	complete	PP-OCRv6_small_rec student trained with hard-label CTC and teacher-logit distillation, evaluated through the adaptive75 two-layer full-page pipeline. Params reports the 5,292,614-parameter deployable recognition architecture; the resumable training state is larger because it retains auxiliary training heads. All 13,338 rows scored with zero missing predictions.
6	Gemma 4 31B IT	31B	35.66	28.6	31.48	13.16	8.05	8.05	93.45	70.34	49.15		Gemma	complete	Provider pinned to cerebras/fp16 with fallback disabled; reasoning disabled.
7	🔒 Qwen 3.6 Flash		37.62	38.63	42.03	16.78	8.95	17.59	67.45	51.21	59.67		Qwen	complete	Reasoning disabled. Alibaba refused 35 images during data inspection; refusals are scored as blank predictions.
8	Chandra OCR 2 NVFP4A16	3B	43.66	36.85	40.76	57.42	49.5	20.27	62.19	46.44	53.43		Chandra	complete	vLLM 0.20.2; batch size 24 per GPU; max new tokens 1536; checkpoint revision b11480ba302227c206eaa4ef558acbacfd7f7eea. All 13,338 rows scored with zero blank predictions.
9	Bina 0.2 Microscopy	1.11M	44.11	48.8	52.36	46.97	18.27	49.2	53.01	30.08	48.39	Paddle Inference · CTC	Bina	complete	PP-OCRv6_tiny_rec student trained with hard-label CTC and teacher-logit distillation, evaluated through the adaptive75 two-layer full-page pipeline. Params reports the 1,113,470-parameter deployable recognition architecture; the resumable training state is larger because it retains auxiliary training heads. All 13,338 rows scored with zero missing predictions.
10	Surya OCR 2	686M	47.94	46.45	48.31	21.51	12.6	18.86	90.71	75.93	74.03		Surya	complete	Stallion RTX 5080 Laptop GPU; batch size 8; max new tokens 1536. All 13,338 rows scored; 64 blank predictions included and penalized.
11	Gemma 4 26B A4B IT	26B	57.2	34.34	37.7	9.99	5.2	9.86	185.39	165.36	58.82		Gemma	complete	Provider pinned to wafer/fp8 with fallback disabled; reasoning disabled.
12	Tesseract Persian (fas)	270K	66.29	69.93	71.04	50.19	34.51	50.91	89	69.56	88.95		Tesseract	complete	Community submission q6qR0L; 897 blank predictions included and penalized. Scored with the frozen Persian normalization protocol. Parameter count covers the 269,887-weight tessdata_fast Persian LSTM; engine code and dictionaries are excluded.
13	DeepSeek-OCR-2	3.39B	68.65	58.61	62.36	35.67	22.74	30.43	156.06	120.38	86.78	vLLM 0.20.1 · BF16 · greedy	DeepSeek	complete	Checkpoint revision aaa02f3811945a91062062994c5c4a3f4c0af2b0; official Free OCR. prompt; 8× RTX PRO 6000 Blackwell 96 GB, one replica per GPU; batch size 64 per GPU; max new tokens 2048; n-gram repetition processor size 40. Dataset revision fefe25215114da8ac0ea21ff5e8f97204b2762d5 was streamed by HTTP Parquet range reads without materializing it. All 13,338 rows scored with zero missing, blank, or failed predictions; 1,844 outputs reached the token cap and were included as generated.
14	WeightedAI Persian OCR	97.4M	76.84	79.54	80.5	58.94	40.12	59.68	100.75	91.3	99.41		PersianOCR	complete	Official OpenCV line segmentation and greedy CTC decoding; line batch size 16; model revision d7edaf2a76e33e77e447bf88df4a2d023bef3b5d. All 13,338 rows scored; latency is the mean over 1,882 saved timing rows.
15	🔒 Gemini 2.5 Flash Lite		104.92	37.53	40.76	217.47	166.86	18.78	235.25	204.43	56.29		Gemini	complete	Provider pinned to google-ai-studio; reasoning excluded from returned transcription.

1

🥇 Bina 0.1 Koochik

666M

9.84

10.21

12.12

10.83

6.54

11.19

11.41

8.38

9.22

Bina · Surya

complete

Persian LoRA checkpoint step 8000, merged for inference; batch size 8 per GPU; max new tokens 4096. All 13,338 rows scored with zero missing predictions. Triple Threat uses the leaderboard's current 20% WER · 20% CER · 60% S³ weighting.

2

🥈 Bina 0.2 Rizeh

35.6M

11.47

11.78

13.42

8.93

2.35

9.85

19.08

13.64

13.7

Paddle Inference · CTC

Bina

complete

PP-OCRv6_medium_rec Stage-8 release with the adaptive75 two-layer full-page pipeline. All 13,338 rows scored with zero missing predictions.

3

🥉 🔒 Gemini 3.5 Flash

15.72

13.97

15.74

6.71

3.68

3.89

33.95

29.06

24.05

Gemini

complete

OpenRouter routing; minimal reasoning effort; reasoning excluded from returned transcription.

4

🔒 Gemini 3.1 Flash Lite

18.62

16.92

18.79

7.22

4.43

6.59

39.5

33.54

27.24

Gemini

complete

OpenRouter routing; default reasoning effort; reasoning excluded from returned transcription.

5

Bina 0.2 RizehPizeh

5.29M

25.17

27.41

30.31

25.22

8.81

27.22

33.13

20.15

27.59

Paddle Inference · CTC

Bina

complete

PP-OCRv6_small_rec student trained with hard-label CTC and teacher-logit distillation, evaluated through the adaptive75 two-layer full-page pipeline. Params reports the 5,292,614-parameter deployable recognition architecture; the resumable training state is larger because it retains auxiliary training heads. All 13,338 rows scored with zero missing predictions.

6

Gemma 4 31B IT

31B

35.66

28.6

31.48

13.16

8.05

93.45

70.34

49.15

Gemma

complete

Provider pinned to cerebras/fp16 with fallback disabled; reasoning disabled.

7

🔒 Qwen 3.6 Flash

37.62

38.63

42.03

16.78

8.95

17.59

67.45

51.21

59.67

Qwen

complete

Reasoning disabled. Alibaba refused 35 images during data inspection; refusals are scored as blank predictions.

8

Chandra OCR 2 NVFP4A16

3B

43.66

36.85

40.76

57.42

49.5

20.27

62.19

46.44

53.43

Chandra

complete

vLLM 0.20.2; batch size 24 per GPU; max new tokens 1536; checkpoint revision b11480ba302227c206eaa4ef558acbacfd7f7eea. All 13,338 rows scored with zero blank predictions.

9

Bina 0.2 Microscopy

1.11M

44.11

48.8

52.36

46.97

18.27

49.2

53.01

30.08

48.39

Paddle Inference · CTC

Bina

complete

PP-OCRv6_tiny_rec student trained with hard-label CTC and teacher-logit distillation, evaluated through the adaptive75 two-layer full-page pipeline. Params reports the 1,113,470-parameter deployable recognition architecture; the resumable training state is larger because it retains auxiliary training heads. All 13,338 rows scored with zero missing predictions.

10

Surya OCR 2

686M

47.94

46.45

48.31

21.51

12.6

18.86

90.71

75.93

74.03

Surya

complete

Stallion RTX 5080 Laptop GPU; batch size 8; max new tokens 1536. All 13,338 rows scored; 64 blank predictions included and penalized.

11

Gemma 4 26B A4B IT

26B

57.2

34.34

37.7

9.99

5.2

9.86

185.39

165.36

58.82

Gemma

complete

Provider pinned to wafer/fp8 with fallback disabled; reasoning disabled.

12

Tesseract Persian (fas)

270K

66.29

69.93

71.04

50.19

34.51

50.91

89

69.56

88.95

Tesseract

complete

Community submission q6qR0L; 897 blank predictions included and penalized. Scored with the frozen Persian normalization protocol. Parameter count covers the 269,887-weight tessdata_fast Persian LSTM; engine code and dictionaries are excluded.

13

DeepSeek-OCR-2

3.39B

68.65

58.61

62.36

35.67

22.74

30.43

156.06

120.38

86.78

vLLM 0.20.1 · BF16 · greedy

DeepSeek

complete

Checkpoint revision aaa02f3811945a91062062994c5c4a3f4c0af2b0; official Free OCR. prompt; 8× RTX PRO 6000 Blackwell 96 GB, one replica per GPU; batch size 64 per GPU; max new tokens 2048; n-gram repetition processor size 40. Dataset revision fefe25215114da8ac0ea21ff5e8f97204b2762d5 was streamed by HTTP Parquet range reads without materializing it. All 13,338 rows scored with zero missing, blank, or failed predictions; 1,844 outputs reached the token cap and were included as generated.

14

WeightedAI Persian OCR

97.4M

76.84

79.54

80.5

58.94

40.12

59.68

100.75

91.3

99.41

PersianOCR

complete

Official OpenCV line segmentation and greedy CTC decoding; line batch size 16; model revision d7edaf2a76e33e77e447bf88df4a2d023bef3b5d. All 13,338 rows scored; latency is the mean over 1,882 saved timing rows.

15

🔒 Gemini 2.5 Flash Lite

104.92

37.53

40.76

217.47

166.86

18.78

235.25

204.43

56.29

Gemini

complete

Provider pinned to google-ai-studio; reasoning excluded from returned transcription.

Persian OCR Triple Threat 🔥 Benchmark

Submit a model