AI repos for PyTorch
Open-source AI projects built around PyTorch or genuinely useful if you work with it, each with what it does for you. 21 repos, most relevant first.
-
LLMs-from-scratch
Build a GPT-style model in plain PyTorch, chapter by chapter, from tokenizer to training loop.
-
minimind
Core training code written in PyTorch from scratch, covering a 64M-parameter model's full pipeline.
-
transformers
Readable PyTorch model definitions for text, vision, audio and multimodal tasks you can customize.
-
reasoning-from-scratch
Notebooks implement reasoning on a Qwen3 base model in PyTorch, step by step.
-
autoresearch
Compact single-GPU LLM training script in PyTorch that the agent edits and re-runs.
-
AI-For-Beginners
Jupyter notebooks teach neural networks in PyTorch alongside TensorFlow with labs and quizzes.
-
Kronos
Run decoder-only Transformer forecasting models in PyTorch with batch prediction and sampling controls.
-
airllm
Run layer-by-layer inference on small GPUs, with optional 4-bit or 8-bit compression.
-
VoxCPM
Run a 2B-parameter text-to-speech model in PyTorch with a Python API, CLI and streaming serving options.
-
timesfm
Run a pretrained forecasting model on a PyTorch backend and get zero-shot quantile forecasts from NumPy arrays.
-
lingbot-map
A PyTorch transformer for streaming 3D reconstruction, with checkpoints, demos and benchmarks to run.
-
tribev2
Multimodal Transformer with pretrained weights and training code for predicting fMRI responses.
-
ComfyUI
PyTorch-based engine that runs Stable Diffusion and other generative models locally on your GPU.
-
whisper
Reference PyTorch implementation of a transformer speech model you can load and run.
-
ai-engineering-from-scratch
Teaches backprop and attention from scratch so you understand what PyTorch does underneath.
-
modular
Write Python model pipelines and fast accelerator kernels in Mojo for ML workloads.
-
InvokeAI
Runs Stable Diffusion and Flux models locally on PyTorch with a canvas and node workflows.
-
Unlimited-OCR
Runs on torch with CUDA, with training and hosting options for the vision-language model.
-
ACE-Step-1.5
A local text-to-music model with LoRA training you can run on CUDA, Mac, AMD or Intel hardware.
-
cosmos
Open world model checkpoints, cookbooks and evaluation suites for video, sound and robot action generation.
-
GLiNER
PyTorch framework for training and running small NER models.