MITvisionRuns locally3.34B (safetensors)
DeepSeek-OCR: Specs, License & How to Run
DeepSeek MIT ~3.3B document OCR. Identification page for Unlimited-OCR-2.
- License
- MIT
- Parameters
- 3.34B (safetensors)
- HF likes
- 3,460 (2026-10-28)
- First seen
- 2025-10-17
TL;DR. DeepSeek-OCR is a vision model from DeepSeek. Parameters: 3.34B (safetensors).License: MIT. It is listed under Adobe Acrobat.
| Developer | DeepSeek |
|---|---|
| Type | vision |
| Architecture | HF model_type deepseek_vl_v2; architecture DeepseekOCRForCausalLM. 12 layers, hidden 1280, vocab 129280, max_position_embeddings 8192. pipeline_tag image-text-to-text. Card title: “DeepSeek-OCR: Contexts Optical Compression.” |
| Parameters | 3.34B (safetensors) |
| Modalities | image, text |
| Languages | multilingual |
| Context | 8,192 tokens (config.json) |
| License | MIT — LICENSE on the Hugging Face card is MIT License (Copyright 2023 DeepSeek). cardData.license is mit (opened 2026-10-28). GitHub deepseek-ai/DeepSeek-OCR is also MIT. Not legal advice. |
| Hardware | Not yet verified — HF usedStorage is 6,711,407,201 bytes (~6.25 GiB). safetensors.parameters sum to 3,336,106,240. The README names Transformers + flash-attn on NVIDIA GPUs. Peak VRAM is not independently measured. (not independently tested) |
| Hugging Face likes | 3,460 (as of 2026-10-28) |
| Model card | https://huggingface.co/deepseek-ai/DeepSeek-OCR |
| First seen | 2025-10-17 |
DeepSeek-OCR is DeepSeek’s ungated ~3.3B document OCR checkpoint. Official hosts we opened on 2026-10-28: the card deepseek-ai/DeepSeek-OCR (created 2025-10-17T06:22:05Z; last modified 2025-11-04T02:36:12Z; 3,460 likes / 2,102,284 downloads), the MIT LICENSE, deepseek-ai/DeepSeek-OCR (24,079 stars / 2,207 forks), and arXiv:2510.18234. API gated was false.
This is an identification page so Unlimited-OCR-2 can cite a same-site OCR peer. It hangs off Adobe Acrobat alternatives. That mapping is editorial. We have no sourced quality comparison against Acrobat or Unlimited-OCR-2. Facts only: Unlimited-OCR-2 vs DeepSeek-OCR.
Hugging Face pipeline_tag is image-text-to-text. The card title is “DeepSeek-OCR: Contexts Optical Compression.” This is not DeepSeek-V4.1-Flash (chat MoE).
What we verified (2026-10-28)
| Field | Value | Source |
|---|---|---|
| Host | huggingface.co/deepseek-ai/DeepSeek-OCR (not gated) | HF API |
| Created / last modified | 2025-10-17T06:22:05Z / 2025-11-04T02:36:12Z | same |
| Likes / downloads | 3,460 / 2,102,284 | same |
| License tag + file | mit / MIT License | HF cardData + LICENSE |
| Pipeline tag | image-text-to-text | HF API |
| Architecture | deepseek_vl_v2 / DeepseekOCRForCausalLM | config.json |
| safetensors.parameters | 3,336,106,240 | HF API |
| usedStorage | 6,711,407,201 bytes (~6.25 GiB) | same |
| Layers / hidden / vocab | 12 / 1280 / 129,280 | config.json |
| Documented context | 8,192 | max_position_embeddings |
| GitHub | deepseek-ai/DeepSeek-OCR · 24,079 stars / 2,207 forks · MIT | GitHub API |
How to run (from the card, not tested)
The README names Transformers (torch==2.6.0, transformers==4.46.3, flash-attn==2.7.3) and deepseek-ai/DeepSeek-OCR with trust_remote_code=True. forkalt has not loaded the checkpoint.
License
LICENSE is MIT License. Card tag mit. GitHub license API also returns MIT. Not legal advice.
Related: Unlimited-OCR-2 — Unlimited-OCR-2 vs DeepSeek-OCR — MiniCPM-OCR 1.0 — MiniCPM-OCR 1.0 vs Unlimited-OCR-2 — Adobe Acrobat alternatives — DeepSeek-V4.1-Flash (chat, not OCR). Catalog: home.
Deploy options
- local: Hugging Face weights (Transformers)
- local: GitHub repo
Comparisons
FAQ
Why is this page on forkalt now?
Unlimited-OCR-2’s card names DeepSeek-OCR as prior work. This is an identification page so that comparison can cite a same-site card. It is not a new DeepSeek release this week.
Is it MIT and downloadable?
Yes. Ungated card deepseek-ai/DeepSeek-OCR (created 2025-10-17T06:22:05Z; 3,460 likes / 2,102,284 downloads on 2026-10-28) tags license mit. The LICENSE file is MIT. GitHub deepseek-ai/DeepSeek-OCR is MIT (24,079 stars / 2,207 forks the same day). Not legal advice.
How big is it?
safetensors.parameters sum to 3,336,106,240 (~3.34B). config.json max_position_embeddings is 8192. The paper title is Contexts Optical Compression (arXiv:2510.18234).
Does it replace Acrobat OCR?
No sourced bake-off. This page identifies an open-weight OCR checkpoint. See Adobe Acrobat alternatives and Unlimited-OCR-2 vs DeepSeek-OCR.
Sources & verification
- https://huggingface.co/deepseek-ai/DeepSeek-OCR
- https://huggingface.co/deepseek-ai/DeepSeek-OCR/raw/main/LICENSE
- https://github.com/deepseek-ai/DeepSeek-OCR
- https://arxiv.org/abs/2510.18234
License notes are not legal advice. Read the original LICENSE. Trademark & disclaimer.