The story so far
Weights, every dated article, and repo commits on one rail.
1 dated article and 30 repo commits so far
Teal is the weights (2024-09-12), amber triangles are 1 dated article (each links to it), grey ticks are repo commits.
GOT-OCR2_0
stepfun-ai Apache 2.0 out 2 years ago
716M parameters, images and text in, text out, 32K tokens of context.
Verified 2026-10-25 Rank #183 in LLMs · score 90.1 method v1 Sources 3 Downloads 30d 297.9k · #219 of 619 tracked
Do people use it?297.9kdownloads, 30 days#219 of 619 we track · top 35%+3 likes, downloads −27% this week1,575 likesHugging Face2026-10-24
Where it's used
Hugging Face does not say who downloads a model or with what. These are the proxies: API providers listing it, quantized builds, fine-tunes/adapters/merges, and Spaces that load it.
No API provider serves it, so its downloads are people running it themselves: 4 quantized builds and 69 Spaces.
- As a service0no API provider listed
- On own hardware4quantized builds
- Built on136 fine-tunes · 7 adapters · 0 merges
- In apps69Hugging Face Spaces
Top Spaces running it: stepfun-ai/GOT_official_online_demo ♥ 361 · Tonic/GOT-OCR ♥ 178 · merve/vision_papers ♥ 118 · tonic-ai/ImageEdit-GOT-OCR ♥ 18
Will it fit?
Weights only, an estimate: leave room for the KV cache. Method
Fits a 16 GB card at INT4
About 0.4 GB at INT4, 0.9 GB at FP8, 1.7 GB at BF16: weights only, so leave room for the KV cache.
Estimate params × bytes per param × 1.2 · how we estimate
What people built on it
17 derivatives on the Hub. Which build to run, by runtime: the most downloaded in each class.
17 derivatives on the Hub, led by 7 adapters
From the Hugging Face model tree; quantized builds in teal, fine-tunes, merges and adapters in greys.
Reference
Everything else, folded: the full tables, every source.
Use-case evidence
4 use cases
What it is good at 4 use cases
Labels are evidence, not verdicts: measured a third-party leaderboard row for this model; author-reported a number the model author published in the card or eval files; structural a deterministic signal from the name, config or chat template, or a decision-model estimate. Each chip links to its source. All use cases.
- Understand images and videostructuraldecision-model estimate p=0.99tag: image-text-to-textpipeline: image-text-to-text
- Extract text from documentsstructuraldecision-model estimate p=1.00name: OCR
- Work across languagesstructuraldecision-model estimate p=0.77tag: multilingual
- Run on a laptop or phonestructuraldecision-model estimate p=0.82fits: edge
Facts
21 fields
Listed in LLMs
Facts
- GOT-OCR2_0 has 716M parameters. (safetensors metadata, observed 2026-10-24)
- Context length is 32,768 tokens (32K). (config.json max_position_embeddings, observed 2026-10-24)
- License: Apache 2.0 (friction: none). Permissive; patent grant. (license file, observed 2026-10-24)
- Created 2024-09-12, last modified 2025-02-04. (HF API, observed 2026-10-24)
| Pipeline tag | image-text-to-text |
|---|---|
| Architectures | GOTQwenForCausalLM |
| Model type | GOT |
| Dtype | bfloat16 |
| Params by dtype | BF16: 716,033,280 |
| Size band | S |
| MoE | no |
| Context length | 32,768 (config.json max_position_embeddings) |
| Languages | multilingual |
| Papers | arXiv:2409.01704, arXiv:2405.14295, arXiv:2312.06109 |
| License | Apache 2.0 (apache-2.0) |
| Hub snapshot | 2026-10-25 |
Servability
vLLM · SGLang
Servability
vLLM: not in registry · 7050ef881ff0835f6890a47db1859a24c65481d6SGLang: not in registry · 773cb938fe9484a8f7e7e46fb1b1fe96412c59ed
vLLM registry (7050ef881ff0835f6890a47db1859a24c65481d6) · SGLang registry (773cb938fe9484a8f7e7e46fb1b1fe96412c59ed) · Matched by architecture class name against each project's registry on main; 'supported' means the architecture is registered, not that this checkpoint was tested.
Hardware fit
3 estimates
Hardware fit ESTIMATE
Weights-only memory for GOT-OCR2_0: params × bytes/param × 1.2 overhead. KV cache and activations are extra. How this is computed.
| BF16 (2 bytes/param) | 1.7 GB · Edge (<4 GB) |
|---|---|
| FP8 (1 byte/param) | 0.9 GB · Edge (<4 GB) |
| INT4 (0.5 bytes/param) | 0.4 GB · Edge (<4 GB) |
Weights only (params × bytes/param × 1.2 overhead); KV cache and activations extra.
All variants
17 variants
Lineage
quantized: 4finetune: 6adapter: 7
Top variants (17 shown, ranked by downloads)
| Variant | Relation | Quant | Params | Band | License | Author | Likes | Downloads |
|---|---|---|---|---|---|---|---|---|
| cstr/got-ocr2-crispembed-GGUF | quantized | GGUF | — | — | apache-2.0 | cstr | 0 | 3.7k |
| pranavdaware/web_ocr | finetune | — | 561M | S | apache-2.0 | pranavdaware | 0 | 487 |
| mlx-community/GOT-OCR2_0-8bit | quantized | -8bit | 561M | S | apache-2.0 | mlx-community | 0 | 341 |
| Solo448/GOT-2-hindi_CPU | finetune | — | 561M | S | apache-2.0 | Solo448 | 1 | 284 |
| mlx-community/GOT-OCR2_0-4bit | quantized | -4bit | 561M | S | apache-2.0 | mlx-community | 0 | 280 |
| mlx-community/GOT-OCR2_0-bf16 | finetune | -bf16 | 561M | S | apache-2.0 | mlx-community | 0 | 232 |
| can-gaa-hou/GOT-OCR2.0-OpenVINO-INT4 | quantized | INT4 | — | — | apache-2.0 | can-gaa-hou | 3 | 0 |
| SHAON123/indicpage-ocr-tamil | adapter | — | — | — | apache-2.0 | SHAON123 | 0 | 98 |
| SHAON123/indicpage-ocr-hindi | adapter | — | — | — | apache-2.0 | SHAON123 | 0 | 96 |
| m4xi/got-mer-lora-r16 | adapter | — | — | — | — | m4xi | 0 | 66 |
| katphlab/got-ocr2-latex | adapter | — | — | — | — | katphlab | 1 | 1 |
| SHAON123/indicpage-ocr-telugu | adapter | — | — | — | apache-2.0 | SHAON123 | 0 | 49 |
| SHAON123/indicpage-ocr-bengali | adapter | — | — | — | apache-2.0 | SHAON123 | 0 | 46 |
| LuoKinging/learning-1 | finetune | — | — | — | mit | LuoKinging | 0 | 0 |
| Effectz-AI/GOT-OCR2_0_Invoice_MD | adapter | — | — | — | — | Effectz-AI | 0 | 0 |
| souvikg544/HindiOCR-VLM | finetune | — | — | — | apache-2.0 | souvikg544 | 0 | 0 |
| Sahil2385/Meenu | finetune | — | — | — | pddl | Sahil2385 | 0 | 0 |
Repository changelog
30 commits
Repository changelog 30 commits
| Date | Commit | SHA |
|---|---|---|
| 2025-02-04 | Update README.md (#40) | 979938b |
| 2024-09-18 | Update modeling_GOT.py | cf6b738 |
| 2024-09-16 | Update README.md | da98dba |
| 2024-09-16 | Update README.md | ce9bff1 |
| 2024-09-14 | Update README.md | 316b0da |
| 2024-09-14 | Update render_tools.py | cd24b73 |
| 2024-09-14 | Update modeling_GOT.py | feafc74 |
| 2024-09-14 | Update README.md | 18a1b8a |
| 2024-09-14 | Update README.md | bc16f65 |
| 2024-09-14 | Update README.md | 75c79d6 |
| 2024-09-14 | Update modeling_GOT.py | ccdbcff |
| 2024-09-14 | Update modeling_GOT.py | 35202c0 |
| 2024-09-14 | Update modeling_GOT.py | 5b7a219 |
| 2024-09-14 | Update modeling_GOT.py | 2035077 |
| 2024-09-13 | Update modeling_GOT.py | 0b6cdba |
| 2024-09-13 | Update README.md | dd35694 |
| 2024-09-13 | Update README.md | cb53b30 |
| 2024-09-13 | Update modeling_GOT.py | ffd1620 |
| 2024-09-13 | Update README.md | d18cb8e |
| 2024-09-13 | Update README.md | c5df14f |
| 2024-09-13 | Update README.md | 41770f3 |
| 2024-09-13 | Update README.md | 43436c3 |
| 2024-09-13 | Update README.md | 42454ed |
| 2024-09-13 | Update README.md | cfcb341 |
| 2024-09-13 | Update README.md | 3cac839 |
| 2024-09-13 | Update README.md | 64c8045 |
| 2024-09-13 | Update README.md | ff68f33 |
| 2024-09-13 | Update README.md | 215794b |
| 2024-09-13 | Update README.md | 2facf01 |
| 2024-09-13 | Update README.md | 11d60b0 |
Articles and coverage
4 pieces
Ranked by source signals (HN, Lobsters, authority domains, official posts) and recency. Summaries are generated from a stored copy of the article text; every sentence cites a span of that text, and nothing is shown that failed the citation check.
Analysis4
github.comauthority957 words
GOT-OCR2.0 is a unified end-to-end OCR model released by UCAS researchers that supports batched inference and fine-tuning. The model has exceeded 1M downloads on Huggingface and is available in an environment using cuda11.8+torch2.0.1. The work is documented in the Arxiv preprint arXiv:2409.01704 and builds upon the Vary architecture and Qwen base model.written, reviewed, cited
- The multi-crop OCR feature does not support batch inference and operates only at the token level.
- Training from stage-1 requires a separate repository, while this codebase only supports post-training stages.
- The ms-swift quick fine-tune command defaults to freezing the vision encoder while training the LLM and projector.
github.comauthority957 words
GOT-OCR2.0 is an open-source OCR model that achieved Huggingface trending #1 status. The project released official codes, weights, and benchmarks in September 2024, with model downloads exceeding 1M on Huggingface. The team operates in an environment using cuda11.8+torch2.0.1 and submitted the paper as arXiv preprint arXiv:2409.01704.written, reviewed, cited
- The multi-crop OCR feature does not support batch inference and operates only at the token level.
- Training from stage-1 requires a separate repository, while this codebase only supports post-training stages.
- The ms-swift quick fine-tune method freezes the vision encoder while training the LLM and projector.
developer.aliyun.com358 words
GOT-OCR2.0 is a unified end-to-end model with 580 million parameters that processes text, formulas, and charts as characters. The training pipeline uses an OPT-125M decoder for efficient pretraining before joint training with a Qwen0.5B decoder.generated, cited
ankitamungalpara.github.io2025-04-051,003 words
GOT-OCR 2.0 is a unified end-to-end model that replaces brittle multi-step OCR pipelines with a single encoder-decoder architecture. The system uses a compact 80M parameter vision encoder and a 0.5B parameter decoder, training with a 6000 token limit before expanding to 8192 tokens for multi-page handling. It employs a dynamic multi-crop strategy for images exceeding 2K resolution and supports outputs including LaTeX, Markdown, and HTML.written, reviewed, cited
- The vision encoder is initialized using VitDet (base) to leverage local attention for high-resolution image processing.
- Post-training expands the decoder's token limit up to 8192 tokens to enable handling of multi-page documents and long outputs.
Papers
3 papers
Papers 3
arXiv links found in the model card and its tags. "Peer-reviewed" is shown only with evidence: a published venue in OpenAlex (measured) or the authors' own journal reference on arXiv (reported).
Similar models
3 models
Similar models
- Florence-2-large · 777M same size band (S), same task (image-text-to-text), permissive license
- surya-ocr-2 · 686M same size band (S), same task (image-text-to-text), more providers
- Florence-2-large-ft · 770M same size band (S), same task (image-text-to-text), permissive license
Sources
3 sources
Sources
Hub snapshot: 2026-10-25 · HF API: 2026-10-24 21:36 UTC · Providers: 2026-10-24 21:44 UTC · Articles: 2026-10-24 21:45 UTC · Summary: 2026-10-07 15:04 UTC · Last verified 2026-10-25
Follow GOT-OCR2_0
Watch this model
- /feeds/models/stepfun-ai--GOT-OCR2_0.xml(RSS 2.0) · every change to GOT-OCR2_0
- /feeds/families/stepfun-ai--GOT-OCR2-0.xml(RSS 2.0) · every change across the GOT-OCR2-0 family
Email alerts are not open yet; the feeds above are the same data. See feeds for the item schema.