.BIN

Introduction

GGML files are the model files of the ggml tensor library from before GGUF. They were used by llama.cpp and other ggml programs in 2023, usually with the extension .bin, and are still the model format of whisper.cpp, whose speech recognition models have names such as ggml-base.en.bin.

The format had no general way to add information: each program wrote a fixed list of settings for its own model type, so new model types and new quantisation methods kept breaking old files. llama.cpp replaced it with GGUF in August 2023 and no longer reads it.

Structure

Numbers are little-endian. The file starts with a 32-bit magic number, written as a number, so the bytes appear reversed:

• ggml (bytes lmgg): the original, unversioned layout
• ggmf (bytes fmgg): followed by a version number
• ggjt (bytes tjgg): followed by a version number, with tensor data aligned to 32 bytes so it can be mapped into memory

Because the hyperparameters have no names, a reader must know in advance which model type wrote the file.

Also used for

• PyTorch model (pytorch_model.bin) .bin

Links