A visual breakdown of six LLM deployment formats used in production: Pickle, Safetensors, GGUF, ONNX, MLX, and TensorRT. Each format trades portability for performance — raw weight formats like Pickle and Safetensors run anywhere Python runs, GGUF bundles everything for framework-free execution, ONNX enables cross-framework portability, MLX is optimized for Apple Silicon's unified memory, and TensorRT compiles to machine code for a specific NVIDIA GPU. The general rule: the more hardware-specific the format, the faster it runs but the fewer machines can use it. The newsletter also covers data leakage in ML pipelines, including common causes (train/test contamination, preprocessing leakage, target-derived features) and prevention strategies.
