Number formats
Dtypes
One canonical name for each kind of number, and the name used inside each published format. Quantized GGUF block types stay on the full reference.
Byte order is not part of the canonical name. It belongs to the file, or to the array when the format lets each array choose. Safetensors is always little-endian. GGUF is little-endian unless the whole version-3 file is big-endian. NumPy, npz, HDF5, and Zarr record the order on the array. PyTorch records it once for the archive. ONNX raw data and Parquet are little-endian. MATLAB marks the file IM for little-endian or MI for big-endian. Text has no byte order.
- bool
- int2NumPyml_dtypes.int2
- uint2NumPyml_dtypes.uint2
- int4NumPyml_dtypes.int4
- uint4NumPyml_dtypes.uint4
- int8
- uint8
- int16
- uint16
- int32
- uint32
- int64
- uint64
- float16
- bfloat16
- float32
- float64
- complex64
- complex128
- float8_e4m3fn
- float8_e5m2
- float8_e4m3fnuz
- float8_e5m2fnuz
- float8_e4m3b11fnuzNumPyml_dtypes.float8_e4m3b11fnuz
- float8_e8m0fnu
- float4_e2m1fn
- float6_e2m3fnsafetensorsF6_E2M3NumPyml_dtypes.float6_e2m3fn
- float6_e3m2fnsafetensorsF6_E3M2NumPyml_dtypes.float6_e3m2fn