Best local OCR and document models in 2026: seven compared, matched to the pages you actually process
TL;DR
There is no single best local OCR model in 2026. For invoices, receipts and standard forms, a compact document VLM such as PaddleOCR-VL or a small Qwen3-VL on one 24 GB GPU is the default.
DeepSeek-OCR is the pick when the job is converting a large scanned archive and throughput per GPU is the constraint; a larger Qwen3-VL or Gemma 4 earns its extra memory on handwriting and hard tables.
Tesseract is the free baseline every model must beat. If a VLM cannot beat Tesseract plus a few regular expressions on your clean invoices, you do not need the VLM for those invoices.
Licenses are not a formality for a European organization: Apache-2.0 and MIT are permissive, the Gemma and Llama 4 terms are not, and the Llama 4 EU clause has surprised more than one procurement team.
This article quotes no benchmark scores on purpose. The only number that predicts your error rate comes from 200 of your own pages, scored per field with one prompt, one schema and one resolution.
Questions people ask
What is the best local OCR model in 2026?
There is no single winner — the right pick follows the page. For invoices, receipts and standard forms, a compact document VLM such as PaddleOCR-VL or a small Qwen3-VL on a 24 GB GPU is the default. DeepSeek-OCR suits large scanned archives where throughput matters most. A larger Qwen3-VL or Gemma 4 handles handwriting and hard tables, and Llama 4 Scout answers questions across a whole contract.
Is Tesseract still good enough for OCR in 2026?
For clean, printed, single-column pages, often yes. Tesseract runs on CPU with no GPU and no prompt, returns text with bounding boxes, and is licensed under Apache-2.0. It does not understand layout, degrades on skewed or photographed pages, and returns nothing useful for handwriting. Its job is to be the floor: run it first, and only pay for a GPU where it falls short on your pages.
Can I run a document OCR model on a single 24 GB GPU?
Yes, for the extraction layer. PaddleOCR-VL is under one billion parameters, DeepSeek-OCR is a few billion with a sparse decoder, and the small Qwen3-VL sizes all fit a single 24 GB card with room for batching. The ~30B Qwen3-VL variants and larger Gemma 4 sizes want 48–80 GB or quantization, and Llama 4 Scout wants an 80 GB GPU with quantization or a multi-GPU box.
Is Mistral OCR a local model?
No. Mistral OCR is an API priced per page; the weights are not published and self-hosting has only been offered through a commercial arrangement. It belongs on a European shortlist because Mistral is an EU-headquartered vendor, which simplifies the GDPR transfer story. If your constraint is that data never leaves your building, it is not the answer unless you have verified the current deployment terms.
Which open-weight OCR models can EU organizations use?
Check the license per version before you download anything. Tesseract, PaddleOCR-VL and the Qwen3-VL sizes we checked ship under Apache-2.0; DeepSeek-OCR under MIT. Gemma 4 uses Google’s own Gemma Terms of Use with use restrictions. The Llama 4 Community License contained, at launch, a clause restricting the multimodal rights of EU-domiciled organizations — have your legal team read the current text.
How do I evaluate OCR models on my own documents?
Build a gold set of about 200 pages stratified by document family, difficulty and language, labeled once against a fixed schema. Give every model identical prompt, schema and image resolution. Score per field after normalization — dates to ISO, amounts to cents — and report the straight-through rate, the share of documents with every required field correct. Measure pages per minute and memory on the card you would actually buy.
Want this worked out on your documents?
We will price a real sample against OpenAI or Anthropic and tell you whether Local document models or a local model on your hardware is the cheaper first step.
There is no single best local OCR model in 2026. There is a short list, and the right pick depends on the page. For the high-volume families most European teams process — invoices, receipts, standard forms — a compact document VLM such as PaddleOCR-VL or a small Qwen3-VL on one 24 GB GPU is the default. DeepSeek-OCR is the pick when the job is converting a large scanned archive and throughput per GPU is the constraint. A larger Qwen3-VL or Gemma 4 earns its extra memory on handwriting and hard tables. Llama 4 Scout is for questions that span a whole 300-page contract. Tesseract is the free baseline every one of them must beat on your pages before you spend money on a GPU. Mistral OCR is a capable API from an EU vendor, not an open-weight model you download.
This article profiles the seven, puts them in one table, and maps them to document families. It quotes no benchmark scores on purpose: public leaderboards are measured on other people’s pages, and the only number that matters is the one you get on yours. The last two sections show how to get that number in a week.
What a “document model” means in 2026
Definition · Document model
A document model is any model that turns a page image or PDF into machine-readable output. In 2026 that covers four layers: classic OCR (characters and boxes), layout analysis (blocks, tables, reading order), vision-language extraction (a schema in, structured JSON out), and long-context language models that answer questions over a whole file. Most production stacks combine two of them.
“OCR” used to mean one thing: detect text, recognize characters, return strings with coordinates. That is still what Tesseract does, and it is still the right tool when pages are clean, printed and single-column. The models below are different in kind, not just better at the old job. It helps to keep the four layers apart:
Classic OCR. Text in, text out. No notion of “invoice total” or “table cell”. Fast on CPU. Fails on skew, stamps, handwriting and dense tables.
Layout analysis. Finds paragraphs, headings, tables and figures and puts them in reading order, usually emitting Markdown or HTML. This is what “document parsing” means in most model cards.
VLM extraction. A vision-language model looks at the page and your field list and returns JSON. One pass replaces OCR, layout and a rules engine. This is where PaddleOCR-VL, DeepSeek-OCR, Qwen3-VL and Gemma 4 sit.
Long-context LLM over a whole file. Feed the parsed text — or many page images — of a full contract or data room into a model with a very long context window and ask questions. This is Llama 4 Scout’s role, and the larger Qwen3-VL sizes can play it too.
A practical stack for a European finance or operations team is two layers: a compact VLM for extraction on the families you see every day, and a long-context model for the occasional whole-file question. Both run on hardware you control, which is what makes on-premise document AI a different category from a cloud OCR API.
The seven models, profiled
Tesseract — the baseline
Tesseract is the open-source OCR engine that has been around for decades, licensed under Apache-2.0, with an LSTM recognizer since version 4 and trained language packs for well over a hundred scripts. It runs on CPU, needs no GPU and no prompt, and produces text with bounding boxes. Its limits are as well known as its strengths: it does not understand layout, it degrades quickly on photographed or skewed pages, and it returns nothing useful for handwriting. Its job in 2026 is to be the floor. If a VLM cannot beat Tesseract plus a few regular expressions on your clean invoices, you do not need the VLM for those invoices.
PaddleOCR-VL is the vision-language successor to the PaddleOCR toolkit, from Baidu’s PaddlePaddle team. It is small — under one billion parameters — and built specifically for document parsing: text, tables, formulas and charts, with reading order preserved. Its model card claims support for over a hundred languages, which matters for teams in the Baltics, the Nordics and Benelux whose pages mix a local language with English. Because it is compact it fits on a single 24 GB card with room for batching, and it is the class of model behind the “20 pages a minute” figure we use for compute cost. The license was Apache-2.0 at release; check the model card for the version you deploy.
DeepSeek-OCR — compression for throughput
DeepSeek-OCR’s idea is to compress a page image into a small number of vision tokens — from a few dozen to a few hundred per page depending on the mode — before a language decoder reads it. Fewer tokens per page means more pages per second on the same GPU, and the paper frames it as a way to convert very large document corpora cheaply. It is a few billion parameters in total with a sparse decoder, so it also fits a 24 GB card. The trade-offs: the aggressive compression modes lose fine detail on dense pages, and its strongest published results are on Chinese and English. The license is MIT at release; verify on the model card. It is the pick when the job is “turn ten years of scanned archive into searchable text” rather than “get every VAT line right”.
Qwen3-VL — the general VLM family
Qwen3-VL from Alibaba’s Qwen team is a family, not a model: sizes run from about two billion parameters to well over two hundred billion, with dense and mixture-of-experts variants and both instruct and reasoning versions. For documents its strengths are tables, charts and following a JSON schema given in the prompt — the extraction workflow rather than plain transcription. The model card lists OCR support for over thirty languages. The small sizes run on 24 GB; the ~30B variants want 48–80 GB or quantization; the largest need several GPUs. Weights ship under Apache-2.0 on the sizes we have checked, but terms have differed by size in earlier Qwen generations, so verify per size before you deploy.
Gemma 4 — Google’s open family
Gemma 4 is the current generation of Google’s open-weight models: multimodal across the family, several sizes, broad multilingual coverage. It is a general model, not a document specialist, so it needs a good prompt and a schema — but it is comfortable with mixed pages, a letter with a table, a form with handwriting, where specialist parsers get brittle. Two things to check before you standardize on it: the Gemma Terms of Use are Google’s own license with use restrictions, not Apache or MIT, and the larger sizes want the same 48–80 GB class as the ~30B Qwen3-VL variants.
Llama 4 Scout — the whole-file model
Llama 4 Scout is Meta’s mixture-of-experts model built around one feature: a context window long enough to hold an entire contract, a full data-room folder, or a year of correspondence in one call. Its active parameter count per token is modest, but total weights are large, so it wants an 80 GB GPU with quantization or a multi-GPU box. Its role in a document stack is not per-page extraction — a compact VLM is faster and cheaper for that — but the second layer: “which of these forty agreements has a change-of-control clause?” For European deployments read the Llama 4 Community License carefully: at launch it contained a clause restricting the rights of EU-domiciled organizations to the multimodal models. Check the current text and your legal team’s reading before you commit.
Mistral OCR — an EU API, not a download
Mistral OCR is Mistral AI’s document model, offered through their API and priced per page rather than per token. It parses text, tables and images into Markdown and handles multilingual pages well. It belongs in this article because European buyers keep asking about it, and because Mistral is an EU-headquartered vendor, which simplifies the GDPR transfer story compared with a US API. It does not belong in the “local” column: the weights are not published, and self-hosting has been offered only through a commercial arrangement. If your constraint is “data never leaves our building”, Mistral OCR is not the answer unless you have verified the current deployment terms. If your constraint is “data stays in the EU and we do not want to run hardware”, it is a serious option.
Side-by-side comparison
Local OCR and document models compared, as of August 2026. Size classes and hardware are approximate; licenses are as published at release — confirm on the model card for the version you deploy. No benchmark scores by design.
Model
Task fit
Size class
License note
Languages
Hardware class
Best for
Tesseract
Classic OCR: text and boxes
Small engine, not an LLM
Apache-2.0
100+ language packs
CPU only
Clean printed pages; the baseline to beat
PaddleOCR-VL
Document parsing VLM: text, tables, formulas, reading order
Under 1B parameters
Apache-2.0 at release; check model card
100+ claimed
Single 24 GB GPU
Multilingual invoices, forms and reports at volume
Whole-file questions over contracts and data rooms
Mistral OCR
API document parser to Markdown
Not published
Proprietary API; self-hosting by arrangement
Multilingual
Mistral’s API (EU vendor)
EU-hosted API without running hardware
Local OCR and document models compared, as of August 2026. Size classes and hardware are approximate; licenses are as published at release — confirm on the model card for the version you deploy. No benchmark scores by design.
Two things the table cannot show. First, the compact models are not “worse” versions of the large ones: on clean, single-family pages a sub-1B document parser can match a 30B general model and run many more pages per hour on the same card. Second, none of these licenses is a formality for a European organization. Apache-2.0 and MIT are permissive. The Gemma and Llama terms are not, and the Llama EU clause in particular has surprised more than one procurement team.
Which model for which document family
Model choice follows the page, and pages come in families. The table below is where we usually start a pilot; the sample set decides where it ends.
Starting shortlist by document family. “First pick” is what we would test first on a fresh sample; “alternative” is the fallback when the first pick misses. Every row assumes a validation layer (totals, VAT arithmetic, date sanity) on top of the model.
Document family
First pick
Alternative
Why
Watch for
Invoices and receipts
PaddleOCR-VL
Qwen3-VL (small)
Compact, fast, layout-aware; a fixed schema suits it
Line-item tables that wrap; photographed receipts
Handwritten forms
Qwen3-VL (mid or large)
Gemma 4
General VLMs read handwriting better than parsers do
Expect a human-review queue; Tesseract is not an option
Contracts and long agreements
PaddleOCR-VL per page + Llama 4 Scout for questions
Large Qwen3-VL with long context
Extraction per page, reasoning over the whole file
License terms for EU use; memory for long inputs
Tables and financial reports
Qwen3-VL
PaddleOCR-VL
Strong table and chart understanding with a schema prompt
Merged cells, footnotes, multi-page tables
Mixed-language pages (e.g. Latvian + English)
PaddleOCR-VL
Gemma 4
Widest language coverage on the card
Verify your exact script and diacritics on a sample
Large scanned archives
DeepSeek-OCR
PaddleOCR-VL
Vision-token compression maximizes pages per GPU-hour
Detail loss in aggressive modes; spot-check dense pages
Starting shortlist by document family. “First pick” is what we would test first on a fresh sample; “alternative” is the fallback when the first pick misses. Every row assumes a validation layer (totals, VAT arithmetic, date sanity) on top of the model.
The pattern across rows: a specialist for volume, a generalist for the mess, and a long-context model only when the question is about the file rather than the page. That is also the architecture behind local document models as we deploy them — the model is one component; the schema and validation around it are the rest.
How to evaluate on your own pages
Leaderboards are measured on public datasets that are cleaner, more English and more uniform than a Latvian accounts-payable inbox. The only evaluation that predicts your error rate is one run on your pages. It does not have to be elaborate:
A gold set. Two hundred pages, stratified by family, difficulty and language, labeled once by a person against a fixed schema. Freeze it and version it.
One prompt, one schema, one resolution. Every model gets identical inputs. If you tune the prompt for one model, tune it for all.
Field-level scoring. Exact match after normalization — dates to ISO, amounts to cents. Report per field, not one blended accuracy.
Straight-through rate. The share of documents where every required field is correct. It is the number finance cares about, because it sets the size of the human queue.
Pages per minute and memory, measured on the card you would actually buy or rent, at the resolution you scored with.
Keep the gold set out of any prompt tuning, and re-run it when a model card changes version. Models in this space update every few months; a shortlist that was right in spring can be wrong by autumn.
A shortlist procedure that takes a week, not a quarter
Teams lose months comparing every model on every page. Narrow first, then measure:
Name the families and the volume. List the document types you process, the monthly pages per type and the languages on them. Rank by pages; most teams find two families make up most of the volume.
Write the schema per family. Field names, types, required or optional, and the tolerance for each: exact, normalized or fuzzy. A model cannot be scored against a schema you have not written.
Run Tesseract first. On clean printed families it may be enough with a few rules, at zero GPU cost. Where it is not, you now know the gap the VLM must close.
Pick two candidates per family from the table above, one compact and one general. Confirm the license is acceptable for an EU organization before you download anything.
Score the gold set with identical prompt, schema and image resolution. Report per-field accuracy, straight-through rate and pages per minute.
Decide by straight-through rate and hardware, not by average accuracy. A model that is slightly less accurate on average but fits a 24 GB card and clears more documents without a human touch is usually the cheaper one to run.
Set a re-test date. Put next quarter’s re-run in the calendar with the gold set attached.
Where the model sits in a European stack
For most European buyers the model is the second decision. The first is where the pages are allowed to go. If supplier invoices, patient claims or HR documents cannot leave the EEA — or you do not want a Schrems II transfer analysis for every new vendor — every open-weight model above runs on hardware in your building or in an EU data center, and the GDPR question shrinks to the one you already answer for your ERP. We covered that in GDPR-compliant document AI. Mistral OCR keeps data in the EU without hardware; the trade is that a third party still processes the file.
The second consideration is cost shape. A compact VLM on a rented A100-class GPU at about €1.80 per hour and roughly 20 pages a minute costs around €0.0015 per page in compute alone, and that figure is flat: page 100,000 costs the same as page one. A per-token API is not. The European buyer’s guide works through how that changes the vendor conversation.
What to do next: pick the two families that make up most of your pages, write their schemas, and run Tesseract, PaddleOCR-VL and one Qwen3-VL size on two hundred real pages. If the straight-through rate on a compact model is good enough, you have your answer and a 24 GB card is your hardware. If it is not, that gold set is what you bring to us — for a model on your hardware, or for bulk document processing on ours.