MITvisionRuns locally3.34B (safetensors)
Unlimited-OCR-2: Specs, License & How to Run
Baidu official MIT ~3.3B long-document OCR (SR-SWA). Not a community fine-tune.
- License
- MIT
- Parameters
- 3.34B (safetensors)
- HF likes
- 228 (2026-10-28)
- First seen
- 2026-10-26
TL;DR. Unlimited-OCR-2 is a vision model from Baidu. Parameters: 3.34B (safetensors).License: MIT. It is listed under Adobe Acrobat.
| Developer | Baidu |
|---|---|
| Type | vision |
| Architecture | HF model_type unlimited-ocr; architecture UnlimitedOCRForCausalLM. language_config: 12 layers, hidden 1280, vocab 129280, 64 routed experts, max_position_embeddings 131072. Card feature: Sparse Reference Sliding Window Attention (SR-SWA). pipeline_tag image-text-to-text. Official Baidu card — not a third-party fine-tune of another OCR checkpoint. |
| Parameters | 3.34B (safetensors) |
| Modalities | image, text |
| Languages | multilingual |
| Context | 131,072 tokens (config.json) |
| License | MIT — Hugging Face cardData.license is mit and Hub tags include license:mit as of 2026-10-28. A raw fetch of resolve/main/LICENSE 404’d (no LICENSE sibling on the card). Not legal advice. |
| Hardware | Not yet verified — HF usedStorage is 7,104,298,077 bytes (~6.62 GiB). safetensors.parameters sum to 3,343,978,497. The README names Transformers (tested on NVIDIA A800-80GB) and a customized SGLang wheel. Peak VRAM is not independently measured. (not independently tested) |
| Hugging Face likes | 228 (as of 2026-10-28) |
| Model card | https://huggingface.co/baidu/Unlimited-OCR-2 |
| First seen | 2026-10-26 |
Unlimited-OCR-2 is Baidu’s ungated ~3.3B long-document OCR checkpoint. It is an official baidu/ card, not a community fine-tune. Official hosts we opened on 2026-10-28: baidu/Unlimited-OCR-2 (created 2026-10-26T03:57:46Z; last modified 2026-10-27T13:35:48Z; 228 likes / 513 downloads) and the paper arXiv:2606.23050 (Unlimited OCR Works). API gated was false. cardData.license is mit; resolve/main/LICENSE 404’d (no LICENSE sibling).
This page hangs off Adobe Acrobat alternatives. That mapping is editorial for document OCR weights next to PDF toolboxes. We have no sourced quality comparison against Acrobat Pro OCR. Same-site OCR peers: Unlimited-OCR-2 vs DeepSeek-OCR, MiniCPM-OCR 1.0 — MiniCPM-OCR 1.0 vs Unlimited-OCR-2.
The README (27 October 2026 release notes) names Sparse Reference Sliding Window Attention (SR-SWA) and “unlimited-horizon” multi-page parsing. Those capability lines stay vendor. Hugging Face pipeline_tag is image-text-to-text. The acknowledgement thanks DeepSeek-OCR, DeepSeek-OCR-2, and PaddleOCR as prior work.
What we verified (2026-10-28)
| Field | Value | Source |
|---|---|---|
| Host | huggingface.co/baidu/Unlimited-OCR-2 (not gated) | HF API |
| Created / last modified | 2026-10-26T03:57:46Z / 2026-10-27T13:35:48Z | same |
| Likes / downloads | 228 / 513 | same |
| License tag + file | mit / no LICENSE sibling (404) | HF cardData + siblings |
| Pipeline tag | image-text-to-text | HF API |
| Architecture | unlimited-ocr / UnlimitedOCRForCausalLM | config.json |
| safetensors.parameters | 3,343,978,497 | HF API |
| usedStorage | 7,104,298,077 bytes (~6.62 GiB) | same |
| Layers / hidden / vocab | 12 / 1280 / 129,280 | config.json language_config |
| Routed experts | 64 | config.json |
| Documented context | 131,072 | max_position_embeddings |
| Paper | arXiv:2606.23050 | card |
How to run (from the card, not tested)
Copied names only. forkalt has not loaded the checkpoint.
Transformers (card: Python 3.10, torch==2.13.0, transformers==4.52.1, NVIDIA A800-80GB): AutoModel.from_pretrained("baidu/Unlimited-OCR-2", trust_remote_code=True). Multi-page: model.infer_multi(...). Single page: model.infer(...).
SGLang is “recommended for serving” on the README and uses a customized wheel named on the card (sglang-0.5.20.dev10+…). That is not stock vLLM.
License
Hub tag license:mit. We did not find a LICENSE blob on the weight repo on 2026-10-28. Treat the grant as the card tag until a file appears. Not legal advice.
What we did not invent
- Card “SOTA” / MarathonDocBench / MPDocBench lines stay vendor.
- Peak VRAM and page-throughput are not independently measured.
- We do not rank Unlimited-OCR-2 against Acrobat, DeepSeek-OCR, or Stirling-PDF’s toolbox OCR.
Related: DeepSeek-OCR — Unlimited-OCR-2 vs DeepSeek-OCR — MiniCPM-OCR 1.0 — MiniCPM-OCR 1.0 vs Unlimited-OCR-2 — Stirling-PDF (PDF toolbox, not this checkpoint) — Stirling-PDF vs Adobe Acrobat. Catalog: home. Editorial: About.
Deploy options
Comparisons
FAQ
Is Unlimited-OCR-2 a random fine-tune?
No. It is Baidu’s ungated official card baidu/Unlimited-OCR-2 (created 2026-10-26T03:57:46Z; 228 likes / 513 downloads on 2026-10-28). The README titles “Unlimited OCR Works 2” and names SR-SWA. The card thanks DeepSeek-OCR as prior work; that is acknowledgement, not a claim that these weights are a community fine-tune.
Is it MIT?
cardData.license is mit and tags include license:mit (opened 2026-10-28). There was no LICENSE file among the card siblings that day (resolve/main/LICENSE 404). Not legal advice.
How big is it?
Hugging Face safetensors.parameters sum to 3,343,978,497 (~3.34B). usedStorage is 7,104,298,077 bytes. config.json sets max_position_embeddings 131072. The card’s “unlimited-horizon” claim is vendor (SR-SWA), not a forkalt measurement.
Does it beat DeepSeek-OCR or Acrobat OCR?
We do not rank it. Card lines about SOTA on MPDocBench / MarathonDocBench stay vendor. See Unlimited-OCR-2 vs DeepSeek-OCR. We have no sourced bake-off against Adobe Acrobat.
Sources & verification
- https://huggingface.co/baidu/Unlimited-OCR-2
- https://arxiv.org/abs/2606.23050
- https://huggingface.co/deepseek-ai/DeepSeek-OCR
License notes are not legal advice. Read the original LICENSE. Trademark & disclaimer.