dots.ocr
RedNote's (小红书) MIT-licensed document parser. A single vision-language model does layout, reading order and text; dots.mocr at 3B added direct SVG output for charts in March 2026.
Overview
hi lab is the research group inside RedNote, better known in China as Xiaohongshu, the social commerce platform. Not an obvious source for a leading document parser, which is part of why the project was slow to register in English-language discussion despite topping benchmarks. The architecture is the point: one vision-language model emits layout regions, reading order and recognised text in a single pass, instead of a detector feeding a recogniser feeding a separate layout and reading-order model. Fewer components means fewer thresholds to tune and fewer places for a pipeline to silently degrade.
The release timeline is worth getting straight because the naming has shifted. dots.ocr shipped on 30 July 2025 at 1.7B parameters. dots.ocr.base, a foundation VLM aimed at OCR pretraining, followed on 31 October 2025. In March 2026 the team rebranded dots.ocr-1.5 as dots.mocr, a 3B model that adds direct conversion of charts and structured graphics into SVG code, with a dots.mocr-svg variant published beside it. Reported OmniDocBench results for dots.ocr are 0.125 overall edit distance in English and 0.160 in Chinese, table TEDS 88.6 and 89.0, and 79.1 on olmOCR-bench. The project's own multilingual benchmark spans roughly a hundred scripts.
Licensing is the cleanest here: plain MIT on code and weights, no user threshold, no revenue trigger, no field-of-use clause. Weights are on Hugging Face, the README is in English and evaluation needs no mainland account of any kind. Practically, budget for a GPU. vLLM has carried official support since 0.11.0 and that is what the team benchmarks against; transformers and CPU inference both run but neither is fast enough for volume. The failure mode is worth designing around too. Any single model that reads and lays out in one pass can produce confident, well-formed text that was never on the page, so a degraded-scan verification step is worth building before this goes anywhere near a compliance or finance workflow.
Key Features
- ✓ One vision-language model produces layout regions, reading order and recognised text in a single pass, with no separate detector stage to tune
- ✓ dots.ocr at 1.7B parameters, dots.mocr at 3B released 19 March 2026, plus a dots.mocr-svg variant published alongside it
- ✓ MIT licence covering both code and weights, with no user cap and no acceptable-use annex
- ✓ Reported OmniDocBench overall edit distance of 0.125 in English and 0.160 in Chinese, with table TEDS at 88.6 and 89.0
- ✓ The project's own dots.ocr-bench spans roughly 100 scripts, with Tibetan, Kannada, Traditional Chinese, Dutch and Russian among the README examples; reported overall edit distance 0.177
- ✓ Official vLLM integration since vLLM 0.11.0, with Hugging Face transformers and CPU inference available as slower fallbacks
Where it holds
- • MIT is as clean as licensing gets in this category. Nothing to negotiate, no threshold to monitor, no clause that changes at scale
- • Collapsing detection, layout and recognition into one model removes an entire class of pipeline tuning: no detector confidence threshold, no box-merging heuristics, no reading-order post-processor
- • Low-resource script coverage is genuinely strong, which is where the older Chinese OCR stacks were weakest
- • Weights are on Hugging Face with an English README and English prompt templates, so nothing here needs a mainland phone number or business licence to evaluate
Where it breaks
- • A single VLM reading and laying out in one pass can hallucinate plausible text on a degraded scan, where a classical detector would simply return nothing; low-quality input needs a verification stage
- • Throughput per page is far below a 34.5M-parameter CNN pipeline, so this is a GPU workload and not a candidate for CPU-only edge deployment
- • Repository and Hugging Face organisation naming has moved between rednote-hilab and dots-studio, which makes pinning one canonical source and one stable model path more annoying than it should be
- • Documentation is thin next to PaddleOCR: a README, a paper and examples, with no fine-tuning recipe of comparable depth
My Take
A social commerce company shipping one of the strongest small document parsers available is not what anyone predicted. The hi lab team inside RedNote put layout detection, reading order and text recognition into a single 1.7B vision-language model under a plain MIT licence, then followed in March 2026 with dots.mocr at 3B, which converts charts and structured graphics straight into SVG. Reported OmniDocBench figures are 0.125 overall edit distance in English and 0.160 in Chinese, with table TEDS at 88.6 and 89.0. Plan on vLLM: transformers inference works but is slow enough that you will not want it anywhere near a batch pipeline.
Quick Info
- Pricing:
- open-source
- Openness:
- Open weights
- Licence:
- MIT
- Starting at:
- Free. MIT licence on code and weights, with no user threshold, no revenue trigger and no field-of-use restriction. Weights are downloadable from Hugging Face and self-hostable. The only cost is GPU time, and in practice that means a card with enough VRAM to hold a 1.7B or 3B vision-language model plus vLLM's KV cache.
- Added:
- Aug 2026
- Updated:
- Aug 2026
Use Cases
Judge it on your own work
The notes above say where dots.ocr holds and where it breaks. The fastest check is your own workload.
Visit website ↗Alternatives to dots.ocr
PaddleOCR
open-sourceBaidu's Apache 2.0 OCR toolkit (飞桨 PaddleOCR). PP-OCRv6 ships at 1.5M, 7.7M and 34.5M parameters, and PaddleOCR-VL-1.6 is a 0.9B document parser covering 109 languages.
MinerU
open-sourceDocument-to-Markdown engine from Shanghai AI Lab's OpenDataLab. Relicensed off AGPLv3 in April 2026, reported at 95.4 on OmniDocBench v1.6 and running on ten domestic AI chip families.
InternVL 3.5
open-sourceShanghai AI Lab's open vision-language family (书生·万象), nine sizes from 1B to 241B-A28B under Apache 2.0. The default open VLM for OCR, document parsing and GUI agent work.
RapidOCR
open-sourcePaddleOCR's models converted to ONNX and shipped without the PaddlePaddle framework. Apache 2.0, a 27.3 MB wheel, PP-OCRv6 small by default and fully offline CPU deployment.