Best open-source OCR and document models in 2026
Document OCR has turned into a race of small models. The models below turn pages, tables, formulas and charts into Markdown or structured text, and most have fewer than 5 billion parameters, so one consumer GPU is enough. Each entry says what the model is built for and what its license asks of you.
-
Unlimited-OCR Most downloaded
Baidu's 3.3B MIT-licensed model built for long documents: it parses dozens to 100+ pages in one pass. It is the most downloaded OCR model here, with about 5.7 million downloads in the 30 days to 11 November 2026.
-
PaddleOCR-VL
Baidu PaddlePaddle's 0.9B Apache 2.0 parser for tables, formulas, charts and seals. It reports 96.33% on OmniDocBench v1.6, a self-reported score that is not comparable with scores on other versions of the benchmark.
-
GLM-OCR
Z.ai's 0.9B MIT-licensed model with an SDK, tables, formulas and form extraction. It reports 94.62 on OmniDocBench V1.5, as reported by the developer.
-
DeepSeek-OCR 2
DeepSeek's 3.4B Apache 2.0 successor to DeepSeek-OCR. Its new encoder reorders visual tokens by meaning instead of reading in scan order.
-
HunyuanOCR
Tencent's 1.1B end-to-end model for parsing, text spotting, information extraction and image translation. Its license does not cover the EU, UK or South Korea.
-
LightOnOCR-3
LightOn's Apache 2.0 family in 4B, 1B and 0.8B sizes. It writes Markdown and can also return bounding boxes, image captions and chart data.
-
Nemotron 3.5 Parse
NVIDIA's 631M page parser returns text, tables, chart data, reading order and boxes, with single, double and triple passes for harder pages. NVIDIA reports 75.20 on ParseBench for the triple pass, only 0.1 above LightOnOCR-3-4B in its own table. OpenMDW-1.1.
-
NuExtract3
NuMind's 4.5B Apache 2.0 model that fills a JSON template you define or writes Markdown, with reasoning you can switch on or off. NuMind's own tests put it ahead of Gemma-4-E4B-it and Qwen3.5 models on its unreleased extraction benchmark, and it had about 40,700 downloads in the 30 days to 11 November 2026.
-
MiniCPM-OCR 1.0 Newest
The newest entry, released on 27 October 2026. A 1.3B Apache 2.0 model for Chinese and English documents with an SDK, a layout stage, and iOS and llama.cpp support.
-
Xiaomi-OCR-0
Xiaomi's 0.9B Apache 2.0 model built on Qwen3.5-0.8B. It ranks first on Real5-OmniDocBench in the developer's own table.
-
Chandra OCR 2
Datalab's 5.3B model for 90+ languages, with Markdown, HTML or JSON output plus layout. Its license is free for research and small teams but bars organizations that offer a competing product.
-
Surya OCR 2
Datalab's 650M model in a 1.4 GB file: text, layout, reading order and tables in 91 languages, run through vLLM or llama.cpp on a GPU, CPU or Apple Silicon. Datalab's own table gives 83.3% on olmOCR-bench, and the weights are free only below US$5 million of revenue and funding.
How we chose
- Open weights you can download today from the developer's official repository.
- Current releases only, so models replaced by a newer version from the same developer, such as the original DeepSeek-OCR, are left out.
- Scores on benchmarks such as OmniDocBench are self-reported and often use different benchmark versions, so we do not rank by them.
- We read each license in full, because some of these models limit commercial use by region or by competing products.