Skip to content
← Models
Open weightsAlibaba X-PLUG (DAMO Academy)/10 trustVerified

mPLUG-DocOwl2

Alibaba's 8B OCR-free document model that compresses each page to 324 tokens — built for multi-page document understanding without a separate OCR step.

Visit official page →

About

DocOwl2 is the second generation of Alibaba's mPLUG document line, and its signature trick is compression: each document page is encoded into just 324 visual tokens, which makes it practical to feed whole multi-page documents to the model instead of one page at a time. It is OCR-free, meaning it reads the page image directly rather than depending on an external text recognizer.

The paper (arXiv 2409.03420, accepted at ACL 2025) claims state-of-the-art results on OCR-free multi-page document understanding at the 8B scale. Code, training data, inference and evaluation code were all released in September 2024, and training code via ms-swift followed that December, so it is genuinely reproducible.

As with most document VLMs, the SOTA claims are author-reported with no independent replication published yet. The 324-token compression is the interesting idea either way: it is a concrete answer to the context-bloat problem that plagues high-resolution document models.

Specs

Context window

324 visual tokens per page; multi-page document input

Pricing

Free / open weights

Free quota

Fully open — no API

API access

No official API — self-host from Hugging Face (mPLUG/DocOwl2) or ModelScope; fine-tune via ms-swift

Weights

Open weights

Released

Not disclosed

Benchmarks

  • Author-reported (arXiv 2409.03420, ACL 2025): SOTA on OCR-free multi-page document understanding at 8B scale (provider-claimed)
  • Each page encoded in 324 tokens vs thousands for typical high-res VLMs (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

document-qamulti-page-documentsocr-free-parsingform-understanding

Trust score — 7/10

/10 trustVerified

Scored 7/10 (TIER verified, metrics-v1). Evidence: peer-reviewed paper (arXiv 2409.03420, accepted ACL 2025) from Alibaba's X-PLUG team, with released code, training data, weights (Apache 2.0), and evaluation code on Hugging Face and ModelScope. Strengths: fully open and reproducible, novel 324-token page compression, multi-page native. Docked for: (1) SOTA claims are author-reported with no independent replication; (2) narrower adoption than the bigger VLM families. License: Apache 2.0.

How the trust score is built

Full methodology →
Adoption · stars/downloads, log-scaled vs peers
6.0/10
Maintenance · recency, releases, issue load
7.0/10
Openness · license + docs
9.0/10
Manual review · sanity read + red-flag check
7.0/10
  • · Peer-reviewed paper (arXiv 2409.03420, accepted ACL 2025) from Alibaba X-PLUG: 8B OCR-free model, each page encoded in 324 tokens, multi-page native.
  • · Released code, training data, inference/eval code, and weights on Hugging Face and ModelScope; fine-tuning via ms-swift.
  • · SOTA claims are author-reported with no independent replication; narrower adoption than the big VLM families.

Sources