For local OCR, a dedicated document model beats a general chat model on most pages: PaddleOCR-VL-1.6 or MinerU fit an 8 GB card, Tesseract or Marker’s CPU mode cover machines with no GPU, and olmOCR-2 or Chandra 2 are the heavier options for 16-24 GB. A general vision LLM such as Qwen3-VL earns its place when the goal is a handful of named fields in JSON (invoice number, total, due date) rather than the whole page as text.
The answer, by job
Sizes are the published weight files in decimal GB, read from the Hugging Face API or the Ollama tag page on 2026-11-16. They are weights only; image tokens and the KV cache need room on top.
| Job | Tool / model | Memory needed | Licence | The catch |
|---|---|---|---|---|
| Clean printed scans, no GPU | Tesseract 5.5.4 | CPU | Apache-2.0 | Text, hOCR, TSV, ALTO or PDF out; table cells and fields are left to you |
| PDFs to Markdown or JSON on a laptop | Marker (--mode fast) or Docling with granite-docling-258M | CPU for Marker fast; granite-docling weights 0.52 GB | Marker code Apache-2.0, weights modified OpenRAIL-M; Docling MIT, granite-docling Apache-2.0 | Marker’s weights carry revenue and funding limits |
| Scans with tables and formulas, 8 GB card | PaddleOCR-VL-1.6 (0.96B) | 1.92 GB weights | Apache-2.0 | Headline score is vendor-reported, and MinerU claims the top spot on the same benchmark |
| Same job, layout-rich JSON | MinerU 4.2.3 with MinerU2.5-Pro-2605 (1.2B) | 2.31 GB weights; MinerU says 8 GB+ VRAM for its GPU tiers | Apache-2.0 plus MinerU terms | Licence thresholds and an attribution duty for online services |
| Multilingual or messy layouts, 12 GB card | dots.mocr (3.0B) or DeepSeek-OCR-2 (3.4B) | 6.08 GB / 6.78 GB weights | MIT / Apache-2.0 | dots card says complex tables and formulas are still hard; DeepSeek’s recipe targets NVIDIA GPUs |
| Bulk PDF to clean text, 16-24 GB card | olmOCR-2-7B-1025-FP8 | 10.06 GB weights; README asks for 12 GB GPU RAM | Apache-2.0 | README asks for an NVIDIA GPU; card says “intended for research and educational use” |
| Handwriting and hard scans, 16-24 GB card | Chandra 2 (5.3B) | 10.59 GB weights (BF16) | Modified OpenRAIL-M | Commercial limits, and the card and LICENSE give different revenue caps |
| Invoice or receipt fields to JSON | Qwen3-VL-8B-Instruct in Ollama with a JSON schema | 6.2 GB (qwen3-vl:8b-instruct) | Apache-2.0 | Generative: it can return a plausible value that is not on the page |
Three kinds of tool
Classic OCR engines find text lines and recognise characters. Tesseract’s README describes its LSTM engine as “focused on line recognition”, with more than 100 languages out of the box. docTR (v1.1.0, Apache-2.0, now maintained by t2k GmbH per its README) and PaddleOCR’s PP-OCRv6 detection and recognition models belong here too. They are fast, run on CPU, and do not invent words, but reading order across columns, table structure and field meaning are your problem.
Document-specialised vision models are small vision-language models trained only to turn a page image into Markdown, HTML or JSON with layout. PaddleOCR-VL-1.6, Surya 2.5, granite-docling-258M, MinerU2.5-Pro, dots.mocr, DeepSeek-OCR-2, olmOCR-2 and Chandra 2 all sit here, mostly between 0.26B and 8.3B parameters. The toolkits around them (PaddleOCR, Surya and Marker, Docling, MinerU, olmOCR) add layout detection, reading order, table reconstruction and batch processing. GOT-OCR 2.0 was an early model of this type; its Hugging Face repositories were last modified in early 2025.
General vision LLMs such as Qwen3-VL read the page and answer instructions. They are the better tool when the question is “what is the total and who issued this?” rather than “give me every word”, because you can hand them a schema. They are the worse tool for transcribing long dense pages, where a free-form decoder has more room to drift. The vision models by VRAM guide covers them in depth.
Picks by hardware
CPU only
Tesseract 5.5.4 (released 2026-10-28) is the baseline for clean printed text. For PDFs that need structure, Marker’s README lists a fast mode that “runs 100% in-process on CPU/MPS (~18 MB Indra ONNX weights)” and a no_ocr mode for born-digital PDFs that uses the embedded text layer. Docling (MIT, v2.150.1) lists “Local execution capabilities for sensitive data and air-gapped environments” and can drive granite-docling-258M from its CLI, whose BF16 weights are 515,093,104 bytes; IBM also publishes a GGUF version. MinerU’s own table puts its basic ONNX tier at about 0.8 GB of downloads and 2 GB of RAM with “None (CPU works)“.
8 GB
PaddleOCR-VL-1.6 is 958,588,736 parameters in a 1.92 GB BF16 file, which leaves about 6.2 GiB of an 8 GiB card free. PaddlePaddle reports 96.33% on OmniDocBench v1.6 (vendor-reported). MinerU2.5-Pro-2605 is 1.2B parameters in 2.31 GB, and its card reports an overall 95.69 on the same benchmark (vendor-reported); both cards claim the top score, so treat neither as settled. Surya 2.5 (689M parameters, 1.38 GB BF16, or an 0.63 GB Q8_0 GGUF plus a 0.21 GB projector) also fits; Datalab reports 83.7% on olmOCR-bench (vendor-reported). For field extraction, qwen3-vl:8b-instruct is 6.2 GB in Ollama, which is tight once a full-page image is in context.
12 GB
dots.mocr, the March 2026 rebrand of dots.ocr-1.5, is 3,039,179,264 parameters in 6.08 GB of BF16 weights (about 6.3 GiB free on a 12 GiB card). DeepSeek-OCR-2 is 3,389,119,360 parameters in 6.78 GB (about 5.7 GiB free); its card gives a Transformers recipe for NVIDIA GPUs with pinned torch==2.6.0, transformers==4.46.3 and flash-attn. The qwen3-vl:8b-instruct-q8_0 tag (9.9 GB) is the higher-precision general option here.
16-24 GB
olmOCR-2-7B-1025-FP8 is 10.06 GB of weights, a fine-tune of Qwen2.5-VL-7B-Instruct. The olmOCR README asks for a recent NVIDIA GPU “with at least 12 GB of GPU RAM” and 30 GB of free disk, so a 16 GB card gives it working room. Ai2 reports 82.4 ± 1.1 on its own olmOCR-Bench for pipeline v0.4.0 (vendor-reported). Chandra 2 is 5.3B parameters in a 10.59 GB BF16 file (about 6.1 GiB free on 16 GiB); Datalab reports 85.8% on olmOCR-bench and describes “Excellent handwriting support” (vendor-reported). At 24 GB, qwen3-vl:30b-a3b-instruct (20 GB) and qwen3-vl:32b-instruct (21 GB) become options for extraction, with little room left for long contexts.
Structured output: getting JSON from an invoice
Most document toolkits emit layout JSON, not business fields. PaddleOCR’s pipelines save_to_json; Surya writes results.json with blocks carrying bbox, polygon, reading_order and confidence; MinerU has mineru parse ... --json; dots.mocr’s prompt asks for “a single JSON object” of layout elements. That tells you where a number is, not that it is the invoice total.
Two documented paths go further. Docling’s DocumentExtractor takes “the desired data schema AKA template, either as a dictionary or as a Pydantic model” and returns the extracted data by page; its example pulls bill_no and total from a Swiss QR-bill image. Ollama’s structured outputs accept a JSON schema in the format field, and the docs state that “Vision models accept the same format parameter”. Marker’s --use_llm flag can “extract values from forms”, but its default service is gemini-3.5-flash, a cloud API; set marker.services.ollama.OllamaService to keep documents local. The structured output guide explains grammar-constrained decoding and what each runtime accepts.
Tables and handwriting
Only vendor claims exist here. PaddleOCR’s README says PP-StructureV3 gives “more fine-grained coordinate information, including table cell coordinates” than the PaddleOCR-VL models. Marker lists --mode vllm for “degraded scans, dense equations, handwriting, & complex tables”. olmOCR lists “Support for equations, tables, handwriting, and complex formatting”. Chandra 2 claims handwriting support and shows handwritten-form examples. The dots.mocr card says “The extraction of complex tables and mathematical formulas persists as a difficult task given the model’s compact architecture.” The local OCR showdown compares several of these tools on public benchmarks.
Licences that restrict commercial use
None of the code licences checked on 2026-11-16 is GPL or AGPL. The restrictions sit elsewhere:
- Surya and Marker code is Apache-2.0, but the weights use a modified AI Pubs Open RAIL-M licence. Its commercial clause bars use if you or your employer “generated more than five million US Dollars ($5,000,000) in gross revenue in the prior year”, raised more than the same amount in equity or debt, or offer a product that “competes with any product or service offered by or made available by Licensor”, except for personal use or research.
- Chandra 2 ships the same LICENSE text with the $5,000,000 figures, while its model card says “startups under $2M funding/revenue”. Read the LICENSE file and ask Datalab if the gap matters to you.
- MinerU is Apache-2.0 “subject to the additional terms”: a separate commercial licence once monthly active users exceed 100 million or monthly revenue exceeds USD 20 million, a duty to state that MinerU is used when offering it as an online service, and automatic termination if either term is broken. The MinerU2.5-Pro weights are tagged apache-2.0 on Hugging Face.
- olmOCR-2 weights are Apache-2.0, with a card note that the model is “intended for research and educational use in accordance with Ai2’s Responsible Use Guidelines”.
- GOT-OCR2.0’s GitHub repository has no LICENSE file (the GitHub licence API returns not found); its README badges say code Apache 2.0 and data CC BY NC 4.0, while the Hugging Face weights are tagged apache-2.0.
- Tesseract, PaddleOCR, PaddleOCR-VL-1.6, docTR, granite-docling, DeepSeek-OCR-2 and Qwen3-VL are Apache-2.0; Docling and dots.mocr are MIT.
Limits
Hallucinated text is the risk that classic OCR does not have. An LLM-based reader generates the most likely next token, so on a blurry digit or a blank region it can write something plausible instead of flagging uncertainty. olmOCR’s own changelog records a release that “fixes auto-rotation detection, and hallucinations on blank documents” (v0.3.0, August 2025). IBM’s granite-docling card says whether smaller models are more susceptible to hallucination “remains uncertain”. Practical guards: keep the classic engine’s confidence scores or Surya’s confidence field, check that line items add up to the total, and route low-confidence pages to a person. Long documents add a second problem covered in the long-document Q&A guide.
Bottom line
Start with the smallest tool that does the job. Tesseract or Marker’s CPU mode for clean text, PaddleOCR-VL-1.6 or MinerU on an 8 GB card for structured pages, dots.mocr or DeepSeek-OCR-2 at 12 GB, and olmOCR-2 or Chandra 2 at 16-24 GB for volume and hard scans. Use a general vision LLM with a JSON schema only for field extraction, and check its numbers. Before any commercial deployment, read the weights licence, not just the code licence: the Datalab and MinerU terms are where the limits are.