Eleven open OCR models in a year: small and hard to compare

Eleven open OCR models arrived between October 2025 and October 2026. Most have 1B to 5B parameters and Apache 2.0 or MIT licenses, but benchmark versions differ, so their scores cannot be ranked.

  • Analysis
  • 3 min read
  • Published

Between 20 October 2025 and 27 October 2026, eleven open OCR and document-parsing models joined our catalog, from DeepSeek-OCR to MiniCPM-OCR. Read together, they show a clear pattern: they are small, most of them carry permissive licenses, and their published scores cannot be lined up against each other.

The eleven models

Model Released Parameters License Standout feature
DeepSeek-OCR (legacy) Oct 2025 3B MIT Compresses page text into few vision tokens
DeepSeek-OCR 2 Jan 2026 3.4B Apache 2.0 An encoder that orders visual tokens by meaning
GLM-OCR Feb 2026 0.9B MIT An SDK and form extraction
Chandra OCR 2 Mar 2026 5.3B Modified OpenRAIL-M 90+ languages and handwriting
PaddleOCR-VL May 2026 0.96B Apache 2.0 Charts and seals
Unlimited-OCR Jun 2026 3.3B MIT Dozens to 100+ pages in one pass
HunyuanOCR Jul 2026 1.1B Tencent Hunyuan Community Translates text inside images
Xiaomi-OCR-0 Sep 2026 0.9B Apache 2.0 Built on Qwen3.5-0.8B
LightOnOCR-3 Oct 2026 0.8B, 1B, 4B Apache 2.0 Bounding boxes and chart data
Nemotron 3.5 Parse Oct 2026 0.63B OpenMDW-1.1 One to three passes trade accuracy for speed
MiniCPM-OCR Oct 2026 1.3B Apache 2.0 Chinese and English, iOS and llama.cpp

They are small

None of the eleven has a main checkpoint above 5.3B parameters, and seven have one of about 1.3B or fewer: GLM-OCR, PaddleOCR-VL, HunyuanOCR, Xiaomi-OCR-0, LightOnOCR-3, Nemotron 3.5 Parse and MiniCPM-OCR. By our estimate, a model of that class needs about 1 to 3 GB of memory in BF16, which fits a consumer GPU or a laptop. A general model such as Qwen3.6, at 27B or 35B, needs far more. Small size is what makes these models practical for bulk jobs such as turning archives of PDFs into text for retrieval-augmented generation.

They are mostly permissive

Nine of the eleven use MIT (three models), Apache 2.0 (five models) or the permissive OpenMDW-1.1 (one model). The two exceptions carry real conditions:

  • Chandra OCR 2 is free for commercial use only for organizations with under US$2 million in prior-year revenue and under US$2 million raised in total, and the license bars products that compete with Datalab.
  • HunyuanOCR is not licensed in the European Union, the United Kingdom or South Korea, and its license forbids using its outputs to improve other AI models.

DeepSeek also switched licenses between generations: the first model is MIT, and DeepSeek-OCR 2 is Apache 2.0.

Downloads tell a different story

Download counts on Hugging Face tell a different story from benchmark tables. Unlimited-OCR had about 5.7 million downloads in the 30 days to 11 November 2026, the most of any OCR model in our catalog, and the first DeepSeek-OCR still had about 1.9 million even though we mark it legacy. Downloads show how many people try a model, not how well it works.

The scores cannot be ranked

Almost every card quotes a score, and the sources differ:

  • GLM-OCR reports 94.62 on OmniDocBench version 1.5, and PaddleOCR-VL reports 96.33% on version 1.6.
  • Xiaomi-OCR-0 reports first place on Real5-OmniDocBench, a derived set, in its own table.
  • Unlimited-OCR reports results on MarathonDocBench, which Baidu created.
  • LightOnOCR-3’s results come from its own blog, which notes that the 4B model scores lower than Infinity Parser Pro on olmOCR-Bench.
  • DeepSeek-OCR 2 publishes no benchmark table on its card.
  • Nemotron 3.5 Parse reports ParseBench, a different benchmark again, with 75.20 in its slowest mode against 75.10 for LightOnOCR-3-4B in NVIDIA’s own table.

Scores from different benchmark versions cannot be compared, and every number above comes from the developer. GLM-OCR vs PaddleOCR-VL shows one such pair.

What to do with this

Choose by feature and license first, and by your own test pages second. Layout boxes and chart data point to LightOnOCR-3, field extraction to GLM-OCR, long multi-page files to Unlimited-OCR, translation of text in images to HunyuanOCR, high-volume parsing on NVIDIA GPUs to Nemotron 3.5 Parse, and phones to MiniCPM-OCR. Our guide how to choose an OCR model turns this into a procedure, and the OCR and document models list keeps the full set in one place.

Sources

Models in this story

Model Released Parameters Context License Smallest variant needs
DeepSeek-OCR DeepSeek 3.34B 570M active – MIT (Permissive) 2.3 GB Laptop or phone
DeepSeek-OCR 2 DeepSeek 3.39B – Apache 2.0 (Permissive) 2.34 GB Laptop or phone
GLM-OCR Z.ai (Zhipu AI) 900M – MIT (Permissive) 0.621 GB Laptop or phone
Chandra OCR 2 Datalab 5.3B 256K OpenRAIL-M (Datalab) (Conditional) 3.66 GB Laptop or phone
PaddleOCR-VL Baidu 960M – Apache 2.0 (Permissive) 0.662 GB Laptop or phone
Unlimited-OCR Baidu 3.34B – MIT (Permissive) 2.3 GB Laptop or phone
HunyuanOCR Tencent 1.12B 128K Tencent Hunyuan (Conditional) 0.773 GB Laptop or phone
Xiaomi-OCR-0 Xiaomi 873M 256K Apache 2.0 (Permissive) 0.602 GB Laptop or phone
LightOnOCR-3 LightOn 853M – 4.54B 256K Apache 2.0 (Permissive) 0.589 GB Laptop or phone
Nemotron 3.5 Parse NVIDIA 631M – OpenMDW-1.1 (Permissive) 0.435 GB Laptop or phone
MiniCPM-OCR 1.0 OpenBMB 1.3B 256K Apache 2.0 (Permissive) 0.897 GB Laptop or phone