timm

What it is and what it does

Timm is a model zoo and training framework for PyTorch image models. It bundles hundreds of pretrained architectures—vision transformers (ViT, EVA, DINOv3), efficient CNNs (EfficientNet, ConvNeXt), and hybrid designs—with their weights hosted on Hugging Face Hub. You use it to load a pretrained model in one line, fine-tune it on your data, or extract features for downstream tasks.

The package depends on torch, torchvision, pyyaml, huggingface_hub, and safetensors. It supports Python 3.8–3.12 and is actively maintained; recent releases add new model families (DINOv3, MobileCLIP-2, SigLIP-2), optimizer variants (Muon, NAdaMuon), and infrastructure improvements (meta-device initialization, pickle security hardening). Training and inference scripts are included.

Use it for:

• Load a pretrained ViT or ConvNeXt model and fine-tune it on your image classification dataset.
• Extract intermediate feature maps from a pretrained model for use in a downstream vision task.
• Benchmark inference speed and accuracy across multiple model architectures on your hardware.
• Train a custom image model from scratch using timm's training scripts and optimizer support.
• Integrate a pretrained encoder (e.g., DINOv3 ViT) into a multimodal or self-supervised pipeline.

Worth the install?

AI-flagged interpretation of the facts on this page — verify before relying

Timm provides a large collection of pretrained PyTorch image models—vision transformers, CNNs, and hybrid architectures—with utilities for loading weights, training, and inference.

Yes. Timm is a mature, actively maintained library with no known vulnerabilities, permissive licensing, and low install friction. It is the de facto standard for accessing a diverse set of pretrained image models in PyTorch. Install it if you need pretrained models, model zoo access, or a training framework for vision tasks.

Install

timm on PyPI

pip install timm

uv add timm

poetry add timm

Installing timm

Before you install

Low friction: pure Python wheel with five runtime dependencies (torch, torchvision, pyyaml, huggingface_hub, safetensors). Active maintenance—last commit 2026-08-11, release 34 days ago—with a large repository (37066 stars) and consistent updates.

License in practice

Apache-2.0 permissive license allows commercial and private use with minimal restrictions; you must include a copy of the license and state significant changes.

Quickstart

pip install timm
import timm
model = timm.create_model('vit_base_patch16', pretrained=True)
output = model(torch.randn(1, 3, 256, 256))

Requires torch and torchvision; model inference requires compatible GPU or CPU with sufficient memory depending on model size.

Verify before relying

• Specific input image dimensions and memory requirements for different model families.
• Compatibility guarantees across PyTorch versions beyond the tested range (1.13 to 2.9.1).
• Whether all model variants and pretrained weights remain available on Hugging Face Hub.

Package facts

Evidence: timm-1.0.28-py3-none-any.whl

Keywords: pytorch, image-classification

Tags

More Software Development packages