# Filetypes

Extension	Description	Saves Training Data?	Security Posture	Common Usage	Used In
.safetensors	A secure and fast tensor storage format designed for storing model weights safely. Prevents arbitrary code execution.	No	High: Prevents arbitrary code execution, more secure than .bin	Model Weights	AI, LLM
.pt	PyTorch's model file format. Can store model architecture, weights, and optimizer states.	No (Usually)	Medium: Potential for arbitrary code execution on load	Model Weights & Training States	AI, LLM, ML
.bin	Generic binary file format used in Hugging Face models (e.g., Transformers). Stores model weights.	No	Medium: May contain arbitrary code execution if mishandled	Model Weights	AI, LLM
.h5	HDF5 format commonly used in Keras and TensorFlow. Stores model architecture and weights.	No	Medium: Generally safe but may store additional code or configs	Model Weights	AI, ML
.onnx	Open Neural Network Exchange format for interoperability between frameworks (e.g., PyTorch to TensorFlow).	No	High: Safe for sharing models across platforms without arbitrary code execution	Model Interoperability	AI, ML
.joblib	Python's Joblib format used for saving scikit-learn models and other Python objects.	Yes	Medium: Can run arbitrary code when loading	Model Weights & Training Data	ML
.pkl	Python’s pickle format for serializing objects, including models.	Yes	Low: Can execute arbitrary code during deserialization	Model Weights & Training Data	ML
.tflite	TensorFlow Lite format for deploying optimized models on mobile and edge devices.	No	High: Optimized for secure and efficient inference on devices	Inference Models	AI, ML
.ckpt	Checkpoint format used in TensorFlow and PyTorch to store model weights and training states.	No (Usually)	Medium: May contain arbitrary code depending on the content	Model Weights & Training States	AI, LLM, ML
.ggml	An optimized format for smaller models used in efficient inference, especially in resource-constrained environments.	No	Medium: Focuses on performance; generally safe for inference	Optimized Inference Models	AI, LLM
.gguf	Successor to .ggml, more optimized for larger models, specifically large language models (LLMs).	No	High: Improved security and efficiency for large models	Optimized Inference Models	AI, LLM