GOT-OCR2.0
580M-parameter unified end-to-end OCR: plain and formatted text, math, tables, charts, sheet music and shapes in one model.
About
GOT-OCR2.0 treats all man-made optical characters, from plain paragraphs to math formulas, tables, charts, sheet music and geometric shapes, as one OCR problem. The 580M-parameter model pairs a high-compression image encoder with a long-context decoder, so it reads whole pages as well as cropped slices and emits either plain text or formatted output such as Markdown, TikZ, SMILES or Kern scores. It also supports interactive region OCR, where you point at a box or a color and it reads just that part, plus multi-page documents and dynamic resolution.
The paper landed in September 2024 and the model has since been folded into Hugging Face Transformers (January 2025), which explains its wide reuse in document pipelines and community demos. For a sub-billion-parameter model it is unusually capable on dense documents, and the formatted-output mode saves a lot of post-processing. Downsides: the training data and evals are author-reported, English and Chinese are the strong languages, and the model is heavy enough that CPU inference is slow. It remains one of the best open starting points for serious document OCR.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- Author-reported results in the paper (General OCR Theory: Towards OCR-2.0); no independent third-party eval found
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 7/10
Scored 7/10 (TIER metrics, metrics-v1). Evidence: paper released 2024-09-03; merged into Hugging Face Transformers 2025-01-31 (HF docs page); official code at github.com/QIN2DIM/GOT-OCR2.0; broad community reuse (many forks, Spaces, tutorials). Strengths: unified OCR-2.0 coverage (text, math, tables, charts, music), formatted output modes, interactive region OCR. Docked for: (1) all benchmark claims are author-reported with no independent eval; (2) license not declared on the checkpoints I could inspect; (3) English/Chinese-centric. License: none declared on model card.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 7.0/10
- Maintenance · recency, releases, issue load
- 6.0/10
- Openness · license + docs
- 7.0/10
- Manual review · sanity read + red-flag check
- 7.0/10
- · Paper released 2024-09-03 (General OCR Theory: Towards OCR-2.0); 580M params, unified end-to-end OCR
- · Merged into Hugging Face Transformers 2025-01-31 (HF docs page)
- · Official code github.com/QIN2DIM/GOT-OCR2.0; weights ucaslcl/GOT-OCR2_0; broad community reuse (forks, Spaces, tutorials)
- · All benchmark claims author-reported; no independent eval found
- · English/Chinese-centric; no license declared on the checkpoints inspected