diff --git a/lib/nnc/ccv_nnc_tensor_io.c b/lib/nnc/ccv_nnc_tensor_io.c
index 112c34c30..9f31dad5b 100644
--- a/lib/nnc/ccv_nnc_tensor_io.c
+++ b/lib/nnc/ccv_nnc_tensor_io.c
@@ -188,7 +188,7 @@ int ccv_nnc_tensor_read(void* const handle, const char* const name, const char*
 	int dim[CCV_NNC_MAX_DIM_ALLOC];
 	memcpy(dim, sqlite3_column_blob(tensor_select_stmt, 4), ccv_min(sizeof(dim), sqlite3_column_bytes(tensor_select_stmt, 4)));
 	const int nd = ccv_nnc_tensor_nd(dim);
-	if (datatype != tensor_params.datatype)
+	if (datatype != tensor_params.datatype && CCV_GET_DATA_TYPE(tensor_params.datatype) != CCV_QX)
 	{
 		// Only ever works for 16F to 32F or 32F to 16F transparently.
 		assert((datatype == CCV_16F && tensor_params.datatype == CCV_32F) || (datatype == CCV_32F && tensor_params.datatype == CCV_16F));
@@ -429,6 +429,9 @@ int ccv_nnc_tensor_read(void* const handle, const char* const name, const char*
 		}
 #endif
 	} else {
+		// If it is QX, we need to have a custom decoder to decode properly.
+		if (datatype != tensor_params.datatype)
+			{ assert(options && options->decode); }
 		size_t data_size = ccv_nnc_tensor_data_size(tensor_params);
 #ifdef HAVE_CUDA
 		if (!options || !options->decode)