MonkeyOCR
A 3B-parameter document foundation model that splits parsing into three questions (where, what, how organized) instead of one giant end-to-end pass.
About
MonkeyOCR reframes document parsing as a Structure-Recognition-Relation triplet: first detect where each block is, then recognize what it says, then predict how the blocks relate. That decomposition lets it avoid both the brittle multi-tool pipelines of older systems and the heavy full-page end-to-end models, and it was trained on MonkeyDoc, a 4.5-million-instance bilingual dataset spanning more than ten document types.
The headline 3B model is joined by smaller CPD (contiguous parameter degradation) variants from 0.6B to 1.2B that trade a small accuracy drop for much faster inference, and the full model is reported to run on a single RTX 3090. The follow-up v2 line adds a document-native vision encoder and separate parsing and understanding checkpoints, so the project is still moving.
The 'beats Gemini 2.5-Pro' headline is author-reported and no independent replication has been published yet, so treat the SOTA claim as provider-claimed. What is solid is the openness: Apache 2.0 code and weights on Hugging Face and ModelScope, plus a public demo.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- Author-reported (arXiv 2506.05218, Jun 2025): 3B model reaches state-of-the-art document parsing, surpassing prior open-source and closed-source methods including Gemini 2.5-Pro (provider-claimed)
- CPD variants (0.6B–1.2B params) keep most of the accuracy at much higher inference speed; full model runs on a single RTX 3090 (author-reported)
- No independent replication published yet — all headline numbers are author-reported
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 8/10
Scored 8/10 (TIER verified, metrics-v1). Evidence: arXiv paper 2506.05218 (submitted 2025-06-05, revised Feb 2026) with named authors, a 4.5M-instance bilingual MonkeyDoc dataset, Apache 2.0 code and weights on Hugging Face and ModelScope, and an actively maintained repo with a v2 encoder/model zoo. Strengths: fully open and reproducible, runs on consumer GPUs, continued development. Docked for: (1) the headline 'beats Gemini 2.5-Pro' claim is author-reported with no independent replication; (2) young project — long-term maintenance unproven. License: Apache 2.0.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 7.0/10
- Maintenance · recency, releases, issue load
- 8.0/10
- Openness · license + docs
- 10.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · arXiv paper 2506.05218 (submitted 2025-06-05, revised Feb 2026): Structure-Recognition-Relation triplet paradigm; 3B doc foundation model on 4.5M-instance bilingual MonkeyDoc.
- · Apache 2.0 code and weights on Hugging Face (echo840/MonkeyOCR) and ModelScope; runs on a single RTX 3090; active v2 encoder/model zoo line.
- · Headline 'beats Gemini 2.5-Pro' claim is author-reported with no independent replication; young project.