Install Docling (2026): pip, uv, CPU-only, GPU & Offline

Q: What is the command to install Docling?

Run pip install docling inside a Python 3.10–3.14 virtual environment. With uv, run uv add docling in a uv project or uv pip install docling in a virtual environment. Check the install with docling --version.

Q: How big is a Docling install?

In our test on Linux, a fresh virtual environment with docling and the CPU build of PyTorch took 1.8 GB on disk, and the default models added 0.5 GB on first run. Using the default PyPI PyTorch on Linux downloads about 3 GB of extra CUDA libraries.

Q: What is the difference between docling and docling-slim?

Since April 2026 the code ships in docling-slim, which installs only the core data model by default and lets you pick features as extras. The docling package is a thin wrapper that installs docling-slim[standard] — the CLI, PDF and Office backends, the layout and table models, RapidOCR and chunking. In our tests with docling 2.148.0, the CLI and all PDF conversion needed PyTorch; since 2.144.0 the Python API can convert Office files with a slim, PyTorch-free install such as docling-slim[format-docx,format-pptx,convert-core]. docling-slim[service-client] is the lightweight option when a docling-serve server does the conversion.

Q: Can I install Docling without internet access?

Yes, if you prepare a machine that has access first: download the wheels with pip download, download the models with docling-tools models download, copy both across, then install from the local wheel folder and point --artifacts-path or DOCLING_ARTIFACTS_PATH at the models folder.

Q: How do I upgrade Docling?

Run pip install -U docling (or uv pip install -U docling) and confirm with docling --version. Docling releases often, so pin a version in production (for example docling==2.148.0) and upgrade deliberately.

Pick your install command#

Choose your operating system, installer, hardware and the features you need. The commands update live and are the same ones we ran on a clean machine.

Requirements at a glance#

Item	Docling 2.148.0	Where we read it
Python	3.10, 3.11, 3.12, 3.13, 3.14 (`>=3.10,<4.0`)	PyPI `requires_python` and classifiers
Operating systems	macOS, Linux, Windows	PyPI classifiers, README
CPU architectures	x86_64 and arm64	README (“both x86_64 and arm64”)
Accelerators	CPU, NVIDIA CUDA, Apple MPS, Intel XPU	`--device` choices: auto, cpu, cuda, mps, xpu
Disk	≈1.8 GB environment + ≈0.5 GB models (CPU PyTorch)	Measured by us with `du -sh`
Python 3.9	Not supported since 2.70.0	README

Step by step with pip#

Create a virtual environment

bash

python3 -m venv .venv
source .venv/bin/activate        # Windows: .venv\Scripts\activate
python -m pip install -U pip

Linux without a GPU? Install CPU-only PyTorch first
The default PyTorch wheel on PyPI for Linux depends on NVIDIA's CUDA libraries. Installing from PyTorch's CPU index first means pip will not fetch them later.
bash
```
pip install torch torchvision --index-url https://download.pytorch.org/whl/cpu
```

Install Docling

bash

pip install docling
# or pin it, which we recommend for production:
pip install "docling==2.148.0"

Verify

bash

docling --version

output on our machine

Docling version: 2.148.0
Docling Core version: 2.106.0
Docling IBM Models version: 4.0.4
Docling Parse version: 7.24.2
Python: cpython-313 (3.13.18)
Platform: Linux-6.18.44-fc-v242-x86_64-with-glibc2.39

Run a first conversion
bash
```
docling convert https://arxiv.org/pdf/2408.09869 --to md --no-ocr
```
The first run downloads the layout model (docling-project/docling-layout-heron) and TableFormer (docling-project/docling-models) from Hugging Face — 506 MB in our cache.

Why the CPU-only step matters#

We read the wheel sizes PyTorch 2.15.0 pulls on Linux x86_64 for Python 3.13. The default build brings cuDNN, cuBLAS, NCCL, Triton and the rest of the CUDA toolkit, even on a server without a GPU.

PyTorch download size on LinuxCompressed wheels pulled for torch 2.15.0 on Linux x86_64, Python 3.13. Docling needs PyTorch for its layout, table and OCR models.

File sizes read from PyPI and download.pytorch.org metadata by docling.org on 6 Nov 2026.

Show data table

PyTorch download size on Linux
	GB
CPU build (download.pytorch.org/whl/cpu)	0.21 GB
Default PyPI build (CUDA 13 dependencies)	2.98 GB

Disk space we measuredInstalled size of a fresh virtual environment (pip) on Linux, Python 3.13, Docling 2.148.0, PyTorch 2.15.0.

Measured with du by docling.org on 6 Nov 2026.

Show data table

Disk space we measured
	GB
docling-slim only (no extras)	0.25 GB
pip install docling + CPU PyTorch	1.79 GB
Default models (first run)	0.51 GB

Installing with uv#

uv resolves Docling's dependencies much faster than pip and can choose the PyTorch build for you with --torch-backend. We confirmed with uv 0.11.32 that --torch-backend cpu resolves torch==2.15.0+cpu.

bashin a uv project

uv init my-app && cd my-app
uv add docling
uv run docling --version

bashin a plain virtual environment

uv venv --python 3.12
source .venv/bin/activate
uv pip install --torch-backend cpu docling     # CPU-only PyTorch
# uv pip install --torch-backend auto docling  # let uv detect your CUDA driver

docling vs docling-slim#

Since 29 April 2026 the code lives in docling-slim, a package that installs only the core data model by default and lets you opt into features. docling is now a thin wrapper that installs docling-slim[standard].

Install	What you get	Use it when
`pip install docling`	CLI, PDF (docling-parse, pypdfium2), Word, PowerPoint, Excel, HTML, Markdown, email, layout + TableFormer models, RapidOCR, chunking, PyTorch	You convert documents on this machine.
`pip install "docling-slim[service-client]"`	The `docling convert-remote` command and the DoclingDocument model — 258 MB environment in our test, no PyTorch	Conversion runs on a docling-serve server.
`pip install docling-slim`	docling-core and the data model only	You only load, inspect or serialise DoclingDocument JSON.

Optional extras#

These are all the extras the docling package declares in 2.148.0, and the packages each one adds.

Extra	Adds	For
`docling[vlm]`	transformers, accelerate, timm, peft, qwen-vl-utils, open-clip-torch; mlx-vlm on Apple Silicon	VLM pipeline (Granite-Docling and other presets), picture descriptions
`docling[asr]`	openai-whisper, whisper-s2t-reborn, numba; mlx-whisper on Apple Silicon	Audio transcription (also needs FFmpeg)
`docling[easyocr]`	easyocr, scikit-image	EasyOCR engine
`docling[tesserocr]`	tesserocr, pandas	Tesseract via its C API (needs system Tesseract)
`docling[rapidocr]`	rapidocr, onnxruntime	RapidOCR on its default onnxruntime backend
`docling[onnxruntime]`	onnxruntime (onnxruntime-gpu where available)	ONNX-based models
`docling[ocrmac]`	ocrmac	Apple Vision OCR on macOS
`docling[htmlrender]`	playwright	Rendering HTML pages before conversion
`docling[xbrl]`	arelle-release	XBRL financial reports
`docling[remote-serving]`	tritonclient[grpc]	Remote model serving (KServe v2 / Triton)

Video files need docling-slim[format-video], which is not exposed as a docling extra. Combine extras in one command: pip install "docling[vlm,easyocr]".

GPU and Apple Silicon#

NVIDIA (Linux, Windows): run nvidia-smi and read the CUDA version it reports — that is the newest CUDA your driver supports. Install PyTorch from a matching index (torch 2.15.0 is published on cu132 and cu134; cu130 and cu126 stop at 2.14.1, which Docling also runs on), then install Docling. Check with python -c "import torch; print(torch.cuda.is_available())".
Apple Silicon: a normal install already includes MPS support. Docling picks the device automatically; force it with --device mps.
Intel GPUs: --device xpu is accepted; you need a PyTorch build with XPU support.

See the performance guide for what a GPU changes and what it does not.

Offline and air-gapped machines#

Docling downloads models from Hugging Face on first use. To run without internet, fetch them in advance on a connected machine:

bashon a machine with internet

pip download "docling==2.148.0" -d wheels/
docling-tools models download -o models/
# default set: layout, TableFormer, code/formula, picture classifier, RapidOCR
# add more by name, e.g.: docling-tools models download granitedocling easyocr -o models/

bashon the offline machine

pip install --no-index --find-links wheels/ docling
export DOCLING_ARTIFACTS_PATH=/path/to/models
docling convert report.pdf --to md
# or per command: docling convert report.pdf --artifacts-path /path/to/models

Without -o, docling-tools models download writes to ~/.cache/docling/models. The DOCLING_ARTIFACTS_PATH variable is read by Docling's settings and used by every pipeline when no explicit path is passed.

Upgrading and pinning#

Docling released 238 versions between July 2024 and November 2026, and flags do change: --force-ocr and --psm are deprecated in 2.148.0 in favour of --ocr-mode full_page and --ocr-option psm=N. Pin the version in production and upgrade on purpose:

bash

pip install -U docling
docling --version
pip freeze | grep -i docling     # record exact versions

Next steps by platform#

Windows

Python from python.org, PATH problems, PowerShell activation and CUDA.

Windows guide →

macOS

Apple Silicon, MPS acceleration and the built-in Vision OCR engine.

macOS guide →

Linux

System libraries, headless servers, CPU-only images and CUDA.

Linux guide →

Docker

The official docling-serve images, their sizes and GPU variants.

Docker guide →

Frequently asked questions

What is the command to install Docling?

Run pip install docling inside a Python 3.10–3.14 virtual environment. With uv, run uv add docling in a uv project or uv pip install docling in a virtual environment. Check the install with docling --version.

How big is a Docling install?

In our test on Linux, a fresh virtual environment with docling and the CPU build of PyTorch took 1.8 GB on disk, and the default models added 0.5 GB on first run. Using the default PyPI PyTorch on Linux downloads about 3 GB of extra CUDA libraries.

What is the difference between docling and docling-slim?

Since April 2026 the code ships in docling-slim, which installs only the core data model by default and lets you pick features as extras. The docling package is a thin wrapper that installs docling-slim[standard] — the CLI, PDF and Office backends, the layout and table models, RapidOCR and chunking. In our tests with docling 2.148.0, the CLI and all PDF conversion needed PyTorch; since 2.144.0 the Python API can convert Office files with a slim, PyTorch-free install such as docling-slim[format-docx,format-pptx,convert-core]. docling-slim[service-client] is the lightweight option when a docling-serve server does the conversion.

Can I install Docling without internet access?

Yes, if you prepare a machine that has access first: download the wheels with pip download, download the models with docling-tools models download, copy both across, then install from the local wheel folder and point --artifacts-path or DOCLING_ARTIFACTS_PATH at the models folder.

How do I upgrade Docling?

Run pip install -U docling (or uv pip install -U docling) and confirm with docling --version. Docling releases often, so pin a version in production (for example docling==2.148.0) and upgrade deliberately.

Sources

Was this accurate? We re-verify pages when Docling releases change behaviour. If a command fails for you, email hello@docling.org with your Docling version and the error — corrections are usually live within a day.

Install Docling the right way — small, fast and offline-ready

Pick your install command#

Requirements at a glance#

Step by step with pip#

Why the CPU-only step matters#

Installing with uv#

docling vs docling-slim#

Optional extras#

GPU and Apple Silicon#

Offline and air-gapped machines#

Upgrading and pinning#

Next steps by platform#

Windows

macOS

Linux

Docker

Frequently asked questions

Sources

Keep going

Windows

macOS

Linux

Docker

Troubleshooting

Config builder