Install

Install Docling on Windows

Windows 10 / 11 (64-bit) with Python 3.10+. The shortest reliable path is 64-bit Python + virtual environment + uv. Verify each command against the official installation docs.

powershell — fastest path

PS>uv add docling

PS>docling --help

Needs 64-bit Python 3.10+ from python.org with “Add to PATH” checked.

uv resolves pre-built wheels and avoids MSVC compiler errors.

Need CUDA? Use WSL2 + Ubuntu — the reliable GPU path.

Requirements

Requirements checklist

Windows 10 or 11, 64-bit (32-bit Python will fail on torch wheels).
Python 3.10, 3.11 or 3.12, 64-bit, installed from python.org with “Add python.exe to PATH” checked.
PowerShell (or Windows Terminal). Commands below assume PowerShell.
~2–4 GB free disk for first-run model downloads.

py --version; py -c "import struct; print(struct.calcsize('P')*8)"

You want 3.10+ and 64. If you see 32, uninstall that Python and install the 64-bit build.

Python

Install 64-bit Python correctly

Download the Windows installer (64-bit) for Python 3.11 or 3.12.
On the first installer screen check “Add python.exe to PATH”, then “Install Now”.
Confirm in a new PowerShell window:

py --version

pip --version

If py is not found, close and reopen the terminal (PATH refresh) or reinstall with the PATH option checked. The Microsoft Store Python shim can also shadow python — prefer the py launcher or the python.org build.

Environment

Virtual environment in PowerShell

py -m venv .venv

.venv\Scripts\Activate.ps1

python -m pip install -U pip

If activation is blocked by the execution policy, run Set-ExecutionPolicy -ExecutionPolicy RemoteSigned -Scope CurrentUser once, then activate again.

Recommended

Fastest install: uv (recommended)

uv resolves pre-compiled wheels and avoids most native-compiler failures on Windows:

uv add docling

With extras (examples):

uv add "docling[rapidocr]"

Or in an existing uv project: uv venv --python 3.12, activate, then uv pip install docling.

Alternative

Standard install: pip

pip install docling

If that succeeds you are done — skip to verification. If pip starts compiling packages and fails with a C++ error, do not fight it: either switch to uv above or fix the toolchain in the next section.

Toolchain fix

Fix “Microsoft Visual C++ 14.0 is required”

Why it happens: some optional dependencies ship native C++ extensions. Without the MSVC toolchain pip falls back to building from source and fails.

Fix A (easiest) — use uv so you get wheels instead of a compiler invocation. Fix B — install the toolchain:

winget install Microsoft.VisualStudio.2022.BuildTools

In the installer select the “Desktop development with C++” workload (MSVC 14.0+).
Restart the terminal, re-activate the venv, delete any half-built env if needed, and retry pip install docling.
Core Docling usually has wheels — this error mostly appears with extras like tesserocr or fasttext. If only an extra needs compiling, consider rapidocr/easyocr instead.

OCR

OCR extras on Windows

Engine	Windows install	Notes
RapidOCR (easiest)	`pip install "docling[rapidocr]"`	Pip-only, no system binary. Good default on Windows.
EasyOCR	`pip install "docling[easyocr]"`	Pip-only, larger download, good multilingual.
Tesseract	Install UB Mannheim binary + set PATH and `TESSDATA_PREFIX`, then `pip install "docling[tesserocr]"`	Best accuracy control, but needs the system binary + language data.

Compare quality and languages in the OCR guide. For scanned PDFs force full-page OCR: docling convert scan.pdf --ocr-mode full_page.

GPU

GPU / CUDA via WSL2

Native Windows CUDA depends on your PyTorch build and drivers. The simplest reliable GPU path is WSL2 with Ubuntu + NVIDIA CUDA drivers for WSL:

Install WSL2 + Ubuntu from the Microsoft Store, install the NVIDIA Windows driver with WSL support.
Inside WSL2 follow the Linux guide (CUDA-enabled torch), then select the device explicitly:

docling convert report.pdf --device cuda

Verify with python -c "import torch; print(torch.cuda.is_available())" — it must print True before GPU flags can help. CPU-only installs are fully supported; GPU only adds speed.

Verification

Verify

docling --help

docling convert sample.pdf --to md

Use a small digital PDF first (no OCR, no GPU). The very first run downloads models — keep the connection alive. Then try --to json and an OCR engine to confirm each layer.

Troubleshooting

Common Windows errors

Installation fails / C++ 14.0 required — see installation fails and MSVC 14.0 required.
python opens the Microsoft Store — use the py launcher or disable the Store alias; prefer the python.org build.
32-bit torch wheels / “not a supported wheel” — you have 32-bit Python; reinstall 64-bit.
Permission errors — never use an admin shell as the fix; a venv plus --user-free install is the fix.
Slow first conversion — normal: models are downloading. Pre-fetch with docling-tools models download --all.

FAQ

Windows FAQ

uv or pip on Windows?

Prefer uv add docling. It resolves pre-compiled wheels and skips the MSVC toolchain that pip sometimes demands for optional extras.

Do I need Visual Studio?

Only if pip tries to compile a native extra. The core package usually installs without it. If you see the C++ 14.0 error, install Build Tools 2022 with the C++ workload via winget, or switch to uv.

Which OCR engine is easiest on Windows?

RapidOCR or EasyOCR — both install with pip alone. Tesseract needs the UB Mannheim system binary plus TESSDATA_PREFIX.

Can I use my NVIDIA GPU on Windows?

Yes. Native CUDA works if your torch build supports it; otherwise WSL2 + Ubuntu with CUDA drivers is the most reliable path, then --device cuda.

PowerShell blocks venv activation?

Run Set-ExecutionPolicy -ExecutionPolicy RemoteSigned -Scope CurrentUser once, then .venv\Scripts\Activate.ps1 again.

OverviewBack to overviewCompare all install options and pick your platform. GPU pathLinux / WSL2CUDA-enabled setup inside WSL2 Ubuntu. NextCommandsConvert PDFs to Markdown, JSON and chunks. Fix errorsTroubleshootingFix MSVC, activation and install failures.

Verified with Docling v2.129.0 · Last checked 2026-09-22 · Official source