Document Parsing Vision-Language Models
Compact vision-language models that read whole pages into text, tables, formulas and charts in reading order.
Tier Grades
Tier A: Frontier AI labs
feasible
Big tech labs release document VLMs.
Tier B: Indonesian AI labs
feasible
Fine-tuning a 0.9B model is within an Indonesian lab budget, as Grab did with 1B and 2B models for Southeast Asian IDs.
Tier C: Institution units such as ITB
feasible
A university unit can fine-tune and benchmark sub-1B models.
Tier D: Organisations such as SEMAI (PoT, PoC, PoV)
feasible
Document parsing feeds RAG and form extraction pilots directly.
Where this comes from
Tiers: a) frontier AI labs, b) Indonesian AI labs, c) institution units such as ITB, d) organisations such as SEMAI delivering proof of technology, concept and value. Each rationale cites the anchors below.
Anchors
- ComputePaddleOCR-VL-0.9B combines a NaViT-style encoder with the ERNIE-4.5-0.3B language model.
- DataSupports 109 languages; training data size is not in the abstract.
- Team size (proxy)Proxy only (author count of an exemplar paper, not a headcount): 18 authors on the PaddleOCR-VL paper (arXiv 2510.14528).
- Time horizonMonths: OmniDocBench leaders change several times a year.
- Shelf-life signalsGeneral multimodal LLMs match dedicated parsers, Indonesian documents enter public benchmarks
- Pipeline roleTier A releases parsers; Tier B fine-tunes them for Indonesian documents; Tier D uses them in extraction and RAG pilots.
Where this comes from
Team size is a proxy (author counts of exemplar papers), not a headcount.
Who and What
Exemplars
Observed actors
Linked models
DeepSeek-OCROCR and documentexemplar
olmOCR (olmOCR-2-7B-1025)OCR and documentexemplar
PaddleOCR-VLOCR and documentexemplar
MinerU2.5 (MinerU-2.5)OCR and documentexemplarUnlimited OCROCR and documentproduced byPaddleOCR-VL-1.6OCR and documentproduced by
Falcon-OCROCR and documentproduced by
GLM-OCROCR and documentproduced byPaddleOCR-VL-1.5OCR and documentproduced byDeepSeek-OCR 2OCR and documentproduced byMinerU2.5-ProOCR and documentproduced by
Nemotron Parse 1.1OCR and documentproduced by
LightOnOCROCR and documentproduced by
dots.ocrOCR and documentproduced by
Granite Vision 3.3 2BOCR and documentproduced by
BaichuanMed OCR 7BOCR and documentproduced byVintern-1B v2OCR and documentproduced by
NougatOCR and documentproduced by
DePlotOCR and documentproduced byPix2StructOCR and documentproduced by
DonutOCR and documentproduced by
Grab custom vision LLM 1BOCR and documentproduced byGrab custom vision LLM 2BOCR and documentproduced by
Mistral OCR 4.1OCR and documentproduced byLinks
R&D TopicDocument Parsing Vision-Language Models
Models 27
Grab custom vision LLM 1BGrab, OCR and documentproduced byGrab custom vision LLM 2BGrab, OCR and documentproduced byPaddleOCR-VL-1.5Baidu, OCR and documentproduced byPaddleOCR-VL-1.6Baidu, OCR and documentproduced byDeepSeek-OCRDeepSeek, OCR and documentexemplarDeepSeek-OCR 2DeepSeek, OCR and documentproduced byDonutNaver, OCR and documentproduced byMinerU2.5 (MinerU-2.5)OpenDataLab, OCR and documentexemplarMinerU2.5-ProOpenDataLab, OCR and documentproduced byNemotron Parse 1.1NVIDIA, OCR and documentproduced byPaddleOCR-VLBaidu, OCR and documentexemplarBaichuanMed OCR 7BBaichuan, OCR and documentproduced by+15
DePlotGoogle, OCR and documentproduced by
Dolphin (document parsing)ByteDance, OCR and documentproduced bydots.ocrrednote (Xiaohongshu), OCR and documentproduced byFalcon-OCRTechnology Innovation Institute (TII), OCR and documentproduced byGLM-OCRZ.ai (Zhipu), OCR and documentproduced byGranite Vision 3.3 2BIBM, OCR and documentproduced by
HunyuanOCRTencent, OCR and documentproduced byLightOnOCRLightOn, OCR and documentproduced byMistral OCR 4.1Mistral AI, OCR and documentproduced by
mPLUG-DocOwl 1.5Alibaba DAMO, OCR and documentproduced byNougatMeta, OCR and documentproduced byolmOCR (olmOCR-2-7B-1025)Allen AI, OCR and documentexemplarPix2StructGoogle, OCR and documentproduced byUnlimited OCRBaidu, OCR and documentproduced byVintern-1B v2OCR and documentproduced byPapers
- OCR-free Document Understanding TransformerGeewook Kim et al., ECCV 2022 (per arXiv comment), 2021 founding
- olmOCR: Unlocking Trillions of Tokens in PDFs with Vision Language ModelsJake Poznanski et al., arXiv preprint, 2025 state of the art
- General OCR Theory: Towards OCR-2.0 via a Unified End-to-end ModelHaoran Wei et al., arXiv preprint, 2024 state of the art
- MinerU: An Open-Source Solution for Precise Document Content ExtractionBin Wang et al., arXiv preprint, 2024 state of the art
- Nougat: Neural Optical Understanding for Academic DocumentsLukas Blecher et al., arXiv preprint, 2023 state of the art
Where this comes from
The papers to read first: where the direction began, its current best, and a survey when one exists.
Sources
- arxiv.org/abs/2510.14528 https://arxiv.org/abs/2510.14528
- How we built a custom vision LLM to improve document processing at Grab engineering.grab.com https://engineering.grab.com/custom-vision-llm-at-grab
To cite this page: Project AURION, AI Use-Cases Research & Innovation, "Document Parsing Vision-Language Models", https://aurion.semai.id/topics/document-parsing-vision-language-models/, records of October 2026.