# AI behind OCR **URL:** \ **Category:** Discussion & Solutions\ **Tags:** solved\ **Created:** [May 13, 2026, 9:21am UTC](https://onehack.st/t/ai-behind-ocr/322001 "2026-05-13T09:21:58Z")\ **Posts on this page:** 5\ **Page:** 1
**Author:** ![Vishal\_Jain2](https://onehack.st/user_avatar/onehack.st/vishal_jain2/32/179340_2.png) [@Vishal\_Jain2](https://onehack.st/u/Vishal_Jain2)\ **Post date:** [May 13, 2026, 9:21am UTC](https://onehack.st/t/ai-behind-ocr/322001/1 "2026-05-13T09:21:58Z")
Who is the ai behind this online ocr tool - [https://textgens.com/ocr-tool/](https://textgens.com/ocr-tool/) I understand that it may be fine-tuned model. Can it be deploy locally on 12gb vram rtx3060? Chances are low but I have to ask anyway … ---
**Author:** ![leaner](https://onehack.st/user_avatar/onehack.st/leaner/32/159863_2.png) [@leaner](https://onehack.st/u/leaner)\ **Post date:** [May 13, 2026, 1:16pm UTC](https://onehack.st/t/ai-behind-ocr/322001/2 "2026-05-13T13:16:08Z")
sir try, deepseek ocr i have used it its good btw here is my tele lets connect computer geek here - Humanloved > **[GitHub - deepseek-ai/DeepSeek-OCR: Contexts Optical Compression](https://github.com/deepseek-ai/DeepSeek-OCR)** > > Contexts Optical Compression ---
**Author:** ![Ze380y](https://onehack.st/user_avatar/onehack.st/ze380y/32/48760_2.png) [@Ze380y](https://onehack.st/u/Ze380y)\ **Post date:** [May 13, 2026, 3:23pm UTC](https://onehack.st/t/ai-behind-ocr/322001/3 "2026-05-13T15:23:18Z")
_Hardware doesn’t say no anymore. The OCR scene flipped six months ago._ * * * **Reading you back:** > **Q1** — Who’s behind textgens? > **Q2** — Is it fine-tuned? > **Q3** — Will it run on a 3060 12GB? And the bigger thing nobody asks out loud: > _“Am I just renting the magic from a website forever, or can I have it sitting in my closet?”_ The answer to that real question is the same as the answer to Q3. * * * ### The three answers, scan-fast | | Answer | | --- | --- | | **Q1** | One-person SaaS wrapper. `hellotextgens@gmail.com`, 7 sister sites in the footer, WordPress, `+07:00` timezone. Almost certainly fronting [**Mistral OCR’s free tier**](https://mistral.ai/news/mistral-ocr-3) or Gemini 2.5 Flash. | | **Q2** | Not a fine-tune — just a wrapper. Nothing to download. What they run lives on Mistral’s servers. The 8-page limit is their throttle, not the API’s. | | **Q3** | **Yes.** And the model that fits is _better_ than what they’re fronting. | * * * ### The flip > **October 2025 dropped six open-source OCR models in one month.** > > The winner is **0.9 billion parameters** — smaller than a basic chat LLM. > > It beats **GPT-5.4, Mistral OCR, and Gemini 2.5 Pro** on accuracy. > > Minimum GPU Baidu officially tested it on: > > **RTX 3060 12GB. Your exact card.** 🎯 * * * ### Meet PaddleOCR-VL-1.5 | | | | --- | --- | | **Model** | [PaddleOCR-VL-1.5-0.9B](https://huggingface.co/PaddlePaddle/PaddleOCR-VL) | | **Min GPU** | RTX 3060 12GB — [Baidu’s FAQ confirms it](https://github.com/PaddlePaddle/PaddleOCR/issues/16823) | | **Score** | 94.5% on OmniDocBench v1.5 — beats every cloud API | | **Languages** | 109 (Hindi, Tamil, Bengali, Punjabi via fine-tune) | | **License** | Apache 2.0 | | **Cost / 1000 pages** | ~$0.09 electricity vs $1 for Mistral API = **167× cheaper** | | **Size** | 0.9B — specialization beat scale | A 0.9B model beating 200B+ general models at one specific task. Doesn’t sound real until you remember Baidu trained it _only_ for documents instead of trying to make it a generalist. * * * ### 🪤 The one trap — and the fix > **Default model-card snippet used vanilla attention.** > > One user’s first run: VRAM ballooned to **45GB on a single page**. Two minutes per page. Looks like the 3060 just died — it didn’t. > > Same user, after switching to **FlashAttention 2** (smarter math, same answer, way less memory): > > ➡ **3.3GB VRAM. 19 seconds per page.** > > Source: [HF discussion #59](https://huggingface.co/PaddlePaddle/PaddleOCR-VL/discussions/59). README’s been patched. Use the current one, not any tutorial older than November 2025. * * * > **Setup that works on a 3060 — copy-paste safe, 15 minutes** > > **Install order matters.** FlashAttention prebuilt wheel BEFORE vLLM — reverse the order and the build breaks every time. Reference: [Alex Zhang’s walkthrough](https://medium.com/@alex_paddleocr/accelerating-paddleocr-vl-with-vllm-via-paddleocr-cli-df7999597a21). > > ```bash > # Fresh venv (don't pollute your global Python) > python -m venv .venv_ocr && source .venv_ocr/bin/activate > > # PaddleOCR with doc-parser > pip install "paddleocr[doc-parser]" > > # FlashAttention prebuilt wheel — THIS before vLLM > pip install https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/download/v0.3.14/flash_attn-2.8.2+cu128torch2.8-cp310-cp310-linux_x86_64.whl > > # vLLM via PaddleOCR's installer > paddleocr install_genai_server_deps vllm > > # Launch — NOTE the 0.3 not 0.8 > paddleocr genai_server \ > --model_name PaddleOCR-VL-1.5-0.9B \ > --backend vllm \ > --port 8118 \ > --backend_config <(echo -e 'gpu-memory-utilization: 0.3') > > ``` > > **Two more if-then signs:** > > > **If it OOMs on launch** → drop `gpu-memory-utilization` from 0.3 to 0.2. The [official FAQ](https://github.com/PaddlePaddle/PaddleOCR/issues/16823) names 0.3 as the 12GB sweet spot, but dense table pages push past it. > > > **If FlashAttention wheel install fails** → wrong build for your stack. The URL above is `cu128 + torch2.8 + python3.10`. Match yours to a different wheel from the [same releases page](https://github.com/mjun0812/flash-attention-prebuild-wheels/releases). > > **Calling the running server from Python:** > > ```python > from paddleocr import PaddleOCRVL > > pipeline = PaddleOCRVL( > vl_rec_backend="vllm-server", > vl_rec_server_url="http://127.0.0.1:8118/v1" > ) > output = pipeline.predict("your_document.pdf") > for res in output: > res.save_to_markdown(save_path="output") > > ``` > > **Three backups if PaddleOCR-VL doesn’t fit your specific workflow:** > > | Model | Why pick this | Link | > | --- | --- | --- | > | **olmOCR-2-7B** | Allen AI, ties GPT-5 & Gemini 2.5 Pro on handwriting | [arxiv](https://arxiv.org/pdf/2510.19817) | > | **DeepSeek-OCR 3B** | Smaller, faster, end-to-end | [HF roundup](https://huggingface.co/blog/ocr-open-models) | > | **Multimodal-OCR UI** | Loads 5 models side-by-side for A/B testing | [repo](https://github.com/PRITHIVSAKTHIUR/Multimodal-OCR) | > > **Honest counter — read before going local:** on [sococrbench](https://noahdasanaike.github.io/posts/sococrbench.html) (independent multilingual handwriting test), GPT-5 and Gemini 2.5 Pro still beat the open-source crowd. PaddleOCR-VL wins on _clean printed text + document structure_. For grandma’s cursive letters or messy mixed-script archives, cloud might still edge ahead. Pick honestly based on YOUR documents. * * * > **be-creative** > > **Bonus loot drop 🎁** > > * * * > > **🪄 Workarounds you can retire today** > > - ~~Hope the cloud free-tier doesn’t get patched~~ → local has no rate limit, no patch risk > - ~~Run Tesseract because it’s “the open-source OCR”~~ → 2026 models are 3-5× more accurate > - ~~Accept the 8-page limit on textgens~~ → local processes the whole batch in one run > - ~~Wait until 24GB GPUs are affordable~~ → 12GB is enough now; the 0.9B fits with room > > * * * > > **🧨 The 167× cost flip — the math** > > | Backend | Price per 1000 pages | > | --- | --- | > | GPT-5.4 vision | ~$15 | > | Mistral OCR API | $1 | > | **Your 3060 + PaddleOCR-VL** | **~$0.09** electricity | > > Source: [CodeSOTA leaderboard economics page](https://www.codesota.com/ocr). > > * * * > > **🔥 The October-2025 wave that flipped the field** > > Six open-source OCR models in one month. Most “best OCR 2025” listicles still recommend Tesseract. > > 1. Nanonets OCR2-3B > 2. **PaddleOCR-VL-0.9B** ← the winner > 3. DeepSeek-OCR-3B > 4. Chandra-OCR-8B > 5. olmOCR-2-7B > 6. LightOnOCR-1B > > All open-source. All fit on consumer GPUs. > > * * * > > **🎮 Side door for Apple Silicon owners reading this** > > Same model deploys via [MLX-VLM](https://github.com/PaddlePaddle/PaddleOCR/issues/16823) on M-series Macs. The 12GB-VRAM constraint becomes “16GB unified memory” on Mac. Same answer, different door. > > * * * > > **🕸 textgens’s sister sites, ranked by parallel-free-quota usefulness** > > If you need free OCR _right now_ while setting up local, the same operator runs six more wrappers on different upstream APIs. Parallel free-quota pools. > > | Sister site | Likely wraps | > | --- | --- | > | [anythingtranslate.com](http://anythingtranslate.com) | Gemini or DeepL | > | [anytranscribe.com](http://anytranscribe.com) | Whisper | > | [anyvoicelab.com](http://anyvoicelab.com) | ElevenLabs free tier | > | onlinechatbot.ai | OpenAI / Claude | > | [dailysimulator.com](http://dailysimulator.com), [fontadvice.com](http://fontadvice.com), [colorany.com](http://colorany.com) | Niche LLM wrappers | * * * **You came in expecting the polite no.** A year ago it would’ve been right. Six open-source models that fit your card overtook the cloud APIs in October 2025. Most “best OCR” listicles haven’t noticed. You said chances were low. Chances aren’t low — they flipped. The 3060 was waiting. 🪄 ---
**Author:** ![Vishal\_Jain2](https://onehack.st/user_avatar/onehack.st/vishal_jain2/32/179340_2.png) [@Vishal\_Jain2](https://onehack.st/u/Vishal_Jain2)\ **Post date:** [May 13, 2026, 5:43pm UTC](https://onehack.st/t/ai-behind-ocr/322001/4 "2026-05-13T17:43:37Z")
i wanted to swear and say f%$#ing thank you … but politely, thank you 🙂 ---
**Author:** ![Abdullah\_Mansoor](https://onehack.st/user_avatar/onehack.st/abdullah_mansoor/32/85009_2.png) [@Abdullah\_Mansoor](https://onehack.st/u/Abdullah_Mansoor)\ **Post date:** [May 13, 2026, 6:37pm UTC](https://onehack.st/t/ai-behind-ocr/322001/5 "2026-05-13T18:37:56Z")
You can also try [OCR space](https://ocr.space/ocrapi). It’s not AI, but their APIs’ free tier offers 25,000 requests per month. However, their PDF upload limit is 3 pages.