**MonkeyOCRv2** takes a position most document-AI releases avoid: it ships the **vision encoder itself** as the product. Rather than handing out one end-to-end OCR pipeline, the HUST-affiliated team behind it releases a standalone, document-native visual backbone that other OCR and document systems can drop in, alongside reference models built on top of it. The code and all model weights are declared **Apache-2.0**, and the repository has collected **1,008 stars and 91 forks** since its **10 July 2026** creation, with commits landing as recently as **19 August 2026**. The accompanying paper is arXiv 2607.11562. The encoder ships in three sizes, and they are unusually small. **MonkeyOCRv2-S** is a 28M-parameter ViT-S and **MonkeyOCRv2-B** a 113M-parameter ViT-B, both pretrained at 1280×28×28 and aimed at recognition, parsing and understanding; **MonkeyOCRv2-AS** is a 21M ViTAEv2-S pretrained at 1760×32×32 for detection and segmentation. Loading one is a two-line `AutoModel.from_pretrained` call against `transformers` with `trust_remote_code=True`, which is the whole point — the backbone is meant to be borrowed, not just benchmarked. What makes the backbone credible is the corpus underneath it. **MonkeyDoc v2** is described by the authors as the largest document-image pretraining image-text pair dataset to date: **113 million document images across 17 languages**, of which **52 million synthetic and 41 million real-world samples** have been released so far, under CC BY 4.0 for the annotations. It is also a genuinely heavy download — the project warns that the processed and compressed dataset needs roughly **10 TB**, and recommends 11 TB of free space before starting. Releasing it at all is the more interesting move: the stated goal is a shared data foundation so document-AI comparisons stop being run on private corpora. On the reference models, the parsing line is the strongest result. **MonkeyOCRv2-B-Parsing** is a 0.7B total-parameter model (0.1B ViT plus a 0.6B LLM) scoring **83.3 overall on MDPBench** across 17 languages — 88.1 on digital-born and 81.7 on photographed documents, 84.2 averaged across nine Latin-script languages and 82.1 across eight non-Latin ones — which the project reports as **first among evaluated open-source models** on the official MDPBench leaderboard as of 22 July 2026. The smaller 0.6B S-Parsing variant trails it by less than a point overall at 82.5. A **DFlash** build released on 24 July enables vLLM serving with up to **2× faster inference**, though it requires vLLM 0.25.1 and CUDA 12.9 or later. The understanding line is more modest: **MonkeyOCRv2-B-Und** at 1.8B parameters scores 57.2 overall, with 79.3 on DocVQA, 64.3 on DT-VQA and 58.1 on OCRBench. Deployment is research-grade rather than turnkey. Parsing runs behind a vLLM service started with `serve.py`, then reached through a CLI, a Gradio web demo or a FastAPI endpoint; there are three separate conda environments in the quick-start, pinned versions for torch, transformers and flash-attn, and a fine-tuning path for both parsing and understanding models. Evaluation coverage extends past parsing to scene text recognition, formula recognition, text detection, document tampering detection and overlapping-text segmentation, with training and evaluation instructions for the last two published on 17 August 2026. The limitations are worth stating plainly. The licence is declared in a **one-line LICENSE file** that points at Apache-2.0 rather than containing its text, so GitHub cannot auto-detect it and downstream tooling will read the repository as unlicensed. The MDPBench ranking is the authors' own reading of a leaderboard they submitted to. MonkeyDoc v2's release is explicitly still underway, image-source terms sit with the dataset card rather than the repo, and the 10 TB footprint puts retraining out of reach for most individuals. And beyond documents, the team is candid that text-rich uses such as scientific papers, historical archives, medical reports and remote-sensing maps are suggestions for the community to try, not tasks they have evaluated.