diff --git a/lib/ccv.h b/lib/ccv.h
index 5e251f8f5..81500a75b 100644
--- a/lib/ccv.h
+++ b/lib/ccv.h
@@ -49,7 +49,7 @@ enum {
 	CCV_64S = 0x08000,
 	CCV_64F = 0x10000,
 	CCV_16F = 0x20000,
-	CCV_8S  = 0x40000, // We can still squeeze in 1 more type, which probably will be 8F. (0xFF000 are for data types).
+	CCV_QX  = 0x40000, // QX is a catch-all for quantized models (anything less than or equal to 1-byte). We can still squeeze in 1 more primitive type, which probably will be 8F or BF16. (0xFF000 are for data types).
 };
 
 enum {
@@ -74,7 +74,7 @@ static const ssize_t _ccv_get_data_type_size[] = {
 #define CCV_MAX_CHANNEL (0xFFF)
 #define CCV_GET_CHANNEL(x) ((x) & 0xFFF)
 #define CCV_GET_STEP(cols, type) (((cols) * CCV_GET_DATA_TYPE_SIZE(type) * CCV_GET_CHANNEL(type) + 3) & -4)
-#define CCV_ALL_DATA_TYPE (CCV_8U | CCV_32S | CCV_32F | CCV_64S | CCV_64F | CCV_16F | CCV_8S)
+#define CCV_ALL_DATA_TYPE (CCV_8U | CCV_32S | CCV_32F | CCV_64S | CCV_64F | CCV_16F | CCV_QX)
 
 enum {
 	CCV_MATRIX_DENSE  = 0x00100000,
diff --git a/lib/nnc/ccv_nnc_cmd.c b/lib/nnc/ccv_nnc_cmd.c
index 2f821679c..eaaf31cb9 100644
--- a/lib/nnc/ccv_nnc_cmd.c
+++ b/lib/nnc/ccv_nnc_cmd.c
@@ -362,10 +362,10 @@ ccv_nnc_cmd_t ccv_nnc_cmd_autotune(const ccv_nnc_cmd_t cmd, const size_t max_wor
 	int tensor_memory = 0, tensor_formats = 0, tensor_datatypes = 0;
 	for (i = 0; i < input_size; i++)
 		if (inputs[i])
-			tensor_memory |= CCV_TENSOR_GET_MEMORY(inputs[i]->info.type), tensor_formats |= inputs[i]->info.format, tensor_datatypes |= inputs[i]->info.datatype;
+			tensor_memory |= CCV_TENSOR_GET_MEMORY(inputs[i]->info.type), tensor_formats |= inputs[i]->info.format, tensor_datatypes |= CCV_GET_DATA_TYPE(inputs[i]->info.datatype);
 	for (i = 0; i < output_size; i++)
 		if (outputs[i])
-			tensor_memory |= CCV_TENSOR_GET_MEMORY(outputs[i]->info.type), tensor_formats |= outputs[i]->info.format, tensor_datatypes |= outputs[i]->info.datatype;
+			tensor_memory |= CCV_TENSOR_GET_MEMORY(outputs[i]->info.type), tensor_formats |= outputs[i]->info.format, tensor_datatypes |= CCV_GET_DATA_TYPE(outputs[i]->info.datatype);
 	// In this case, we cannot determine the type of the tensor, skip auto-tune.
 	if (!tensor_memory)
 		return cmd;
@@ -682,10 +682,10 @@ int ccv_nnc_cmd_exec(const ccv_nnc_cmd_t cmd, const ccv_nnc_hint_t hint, const i
 		int tensor_memory = 0, tensor_formats = 0, tensor_datatypes = 0;
 		for (i = 0; i < input_size; i++)
 			if (inputs[i])
-				tensor_memory |= CCV_TENSOR_GET_MEMORY(inputs[i]->info.type), tensor_formats |= inputs[i]->info.format, tensor_datatypes |= inputs[i]->info.datatype;
+				tensor_memory |= CCV_TENSOR_GET_MEMORY(inputs[i]->info.type), tensor_formats |= inputs[i]->info.format, tensor_datatypes |= CCV_GET_DATA_TYPE(inputs[i]->info.datatype);
 		for (i = 0; i < output_size; i++)
 			if (outputs[i])
-				tensor_memory |= CCV_TENSOR_GET_MEMORY(outputs[i]->info.type), tensor_formats |= outputs[i]->info.format, tensor_datatypes |= outputs[i]->info.datatype;
+				tensor_memory |= CCV_TENSOR_GET_MEMORY(outputs[i]->info.type), tensor_formats |= outputs[i]->info.format, tensor_datatypes |= CCV_GET_DATA_TYPE(outputs[i]->info.datatype);
 		backend = ccv_nnc_cmd_find_backend(cmd, tensor_memory, tensor_formats, tensor_datatypes);
 	}
 	assert(backend != CCV_NNC_NO_BACKEND);
diff --git a/lib/nnc/ccv_nnc_graph.c b/lib/nnc/ccv_nnc_graph.c
index dd3679a8e..2c8f6ee20 100644
--- a/lib/nnc/ccv_nnc_graph.c
+++ b/lib/nnc/ccv_nnc_graph.c
@@ -352,7 +352,7 @@ void ccv_nnc_graph_exec_set_io(ccv_nnc_graph_t* const graph, const ccv_nnc_graph
 		if (info->inputs[i])
 		{
 			ccv_nnc_tensor_t* const tensor = CCV_IS_TENSOR_MULTIVIEW(info->inputs[i]) ? _ccv_nnc_any_tensor_from_tensor_multiview((ccv_nnc_tensor_multiview_t*)info->inputs[i]) : info->inputs[i];
-			tensor_memory |= CCV_TENSOR_GET_MEMORY(tensor->info.type), tensor_formats |= tensor->info.format, tensor_datatypes |= tensor->info.datatype;
+			tensor_memory |= CCV_TENSOR_GET_MEMORY(tensor->info.type), tensor_formats |= tensor->info.format, tensor_datatypes |= CCV_GET_DATA_TYPE(tensor->info.datatype);
 		}
 	info->cmd.backend = ccv_nnc_cmd_find_backend(info->cmd, tensor_memory, tensor_formats, tensor_datatypes);
 	info->input_size = input_size;
@@ -415,7 +415,7 @@ ccv_nnc_graph_exec_t ccv_nnc_graph_exec_new(ccv_nnc_graph_t* const graph, const
 			if (info.inputs[i])
 			{
 				ccv_nnc_tensor_t* const tensor = CCV_IS_TENSOR_MULTIVIEW(info.inputs[i]) ? _ccv_nnc_any_tensor_from_tensor_multiview((ccv_nnc_tensor_multiview_t*)info.inputs[i]) : info.inputs[i];
-				tensor_memory |= CCV_TENSOR_GET_MEMORY(tensor->info.type), tensor_formats |= tensor->info.format, tensor_datatypes |= tensor->info.datatype;
+				tensor_memory |= CCV_TENSOR_GET_MEMORY(tensor->info.type), tensor_formats |= tensor->info.format, tensor_datatypes |= CCV_GET_DATA_TYPE(tensor->info.datatype);
 			}
 		info.cmd.backend = ccv_nnc_cmd_find_backend(info.cmd, tensor_memory, tensor_formats, tensor_datatypes);
 	}
diff --git a/lib/nnc/ccv_nnc_symbolic_graph.c b/lib/nnc/ccv_nnc_symbolic_graph.c
index 990be6c51..5dfdf4ec6 100644
--- a/lib/nnc/ccv_nnc_symbolic_graph.c
+++ b/lib/nnc/ccv_nnc_symbolic_graph.c
@@ -838,7 +838,7 @@ static void _ccv_nnc_graph_exec_symbol_set_io(ccv_nnc_symbolic_graph_t* const gr
 		{
 			const ccv_nnc_tensor_symbol_info_t* const tensor_info = (ccv_nnc_tensor_symbol_info_t*)ccv_array_get(graph->tensor_symbol_info, d);
 			tensor_auto = tensor_auto || ccv_nnc_is_tensor_auto(tensor_info->info);
-			tensor_memory |= CCV_TENSOR_GET_MEMORY(tensor_info->info.type), tensor_formats |= tensor_info->info.format, tensor_datatypes |= tensor_info->info.datatype;
+			tensor_memory |= CCV_TENSOR_GET_MEMORY(tensor_info->info.type), tensor_formats |= tensor_info->info.format, tensor_datatypes |= CCV_GET_DATA_TYPE(tensor_info->info.datatype);
 		}
 	}
 	for (i = 0; i < output_size; i++)
@@ -849,7 +849,7 @@ static void _ccv_nnc_graph_exec_symbol_set_io(ccv_nnc_symbolic_graph_t* const gr
 		{
 			const ccv_nnc_tensor_symbol_info_t* const tensor_info = (ccv_nnc_tensor_symbol_info_t*)ccv_array_get(graph->tensor_symbol_info, d);
 			tensor_auto = tensor_auto || ccv_nnc_is_tensor_auto(tensor_info->info);
-			tensor_memory |= CCV_TENSOR_GET_MEMORY(tensor_info->info.type), tensor_formats |= tensor_info->info.format, tensor_datatypes |= tensor_info->info.datatype;
+			tensor_memory |= CCV_TENSOR_GET_MEMORY(tensor_info->info.type), tensor_formats |= tensor_info->info.format, tensor_datatypes |= CCV_GET_DATA_TYPE(tensor_info->info.datatype);
 		}
 	}
 	// If there is no auto tensor, we try to find backend (we don't know which backend if the tensor is auto).
diff --git a/lib/nnc/ccv_nnc_tensor.c b/lib/nnc/ccv_nnc_tensor.c
index cac541ba1..7963c8364 100644
--- a/lib/nnc/ccv_nnc_tensor.c
+++ b/lib/nnc/ccv_nnc_tensor.c
@@ -26,11 +26,11 @@ ccv_nnc_tensor_t* ccv_nnc_tensor_new(const void* const ptr, const ccv_nnc_tensor
 		tensor->info = params;
 		if (tfb)
 		{
-			tensor->type = CCV_NO_DATA_ALLOC | CCV_MATRIX_DENSE | params.datatype | params.dim[2];
+			tensor->type = CCV_NO_DATA_ALLOC | CCV_MATRIX_DENSE | CCV_GET_DATA_TYPE(params.datatype) | params.dim[2];
 			// This corresponding to mat->step
-			tensor->info.dim[4] = CCV_GET_STEP(params.dim[1], (params.datatype | params.dim[2]));
+			tensor->info.dim[4] = CCV_GET_STEP(params.dim[1], (CCV_GET_DATA_TYPE(params.datatype) | params.dim[2]));
 		} else // This won't be recognized by ccv_dense_matrix_t
-			tensor->type = CCV_NO_DATA_ALLOC | CCV_MATRIX_DENSE | params.datatype;
+			tensor->type = CCV_NO_DATA_ALLOC | CCV_MATRIX_DENSE | CCV_GET_DATA_TYPE(params.datatype);
 		tensor->data.u8 = (uint8_t*)ptr;
 		return tensor;
 	}
@@ -77,11 +77,11 @@ ccv_nnc_tensor_t* ccv_nnc_tensor_new(const void* const ptr, const ccv_nnc_tensor
 	tensor->info = params;
 	if (tfb)
 	{
-		tensor->type = CCV_UNMANAGED | CCV_MATRIX_DENSE | params.datatype | params.dim[2];
+		tensor->type = CCV_UNMANAGED | CCV_MATRIX_DENSE | CCV_GET_DATA_TYPE(params.datatype) | params.dim[2];
 		// This corresponding to mat->step
-		tensor->info.dim[4] = CCV_GET_STEP(params.dim[1], (params.datatype | params.dim[2]));
+		tensor->info.dim[4] = CCV_GET_STEP(params.dim[1], (CCV_GET_DATA_TYPE(params.datatype) | params.dim[2]));
 	} else
-		tensor->type = CCV_UNMANAGED | CCV_MATRIX_DENSE | params.datatype;
+		tensor->type = CCV_UNMANAGED | CCV_MATRIX_DENSE | CCV_GET_DATA_TYPE(params.datatype);
 	return tensor;
 }
 
@@ -100,11 +100,11 @@ ccv_nnc_tensor_t* ccv_nnc_tensor_resize(ccv_nnc_tensor_t* const tensor, const cc
 #endif
 	if (tfb)
 	{
-		tensor->type = CCV_UNMANAGED | CCV_MATRIX_DENSE | params.datatype | params.dim[2];
+		tensor->type = CCV_UNMANAGED | CCV_MATRIX_DENSE | CCV_GET_DATA_TYPE(params.datatype) | params.dim[2];
 		// This corresponding to mat->step
-		tensor->info.dim[4] = CCV_GET_STEP(params.dim[1], (params.datatype | params.dim[2]));
+		tensor->info.dim[4] = CCV_GET_STEP(params.dim[1], (CCV_GET_DATA_TYPE(params.datatype) | params.dim[2]));
 	} else
-		tensor->type = CCV_UNMANAGED | CCV_MATRIX_DENSE | params.datatype;
+		tensor->type = CCV_UNMANAGED | CCV_MATRIX_DENSE | CCV_GET_DATA_TYPE(params.datatype);
 	if (size <= tensor->data_size) // Nothing.
 	{
 #ifdef HAVE_CUDA
@@ -177,11 +177,11 @@ ccv_nnc_tensor_t ccv_nnc_tensor(const void* const ptr, const ccv_nnc_tensor_para
 	}
 	if (tfb)
 	{
-		tensor.type = CCV_NO_DATA_ALLOC | CCV_UNMANAGED | CCV_MATRIX_DENSE | params.datatype | params.dim[2];
+		tensor.type = CCV_NO_DATA_ALLOC | CCV_UNMANAGED | CCV_MATRIX_DENSE | CCV_GET_DATA_TYPE(params.datatype) | params.dim[2];
 		// This corresponding to mat->step
-		tensor.info.dim[4] = CCV_GET_STEP(params.dim[1], (params.datatype | params.dim[2]));
+		tensor.info.dim[4] = CCV_GET_STEP(params.dim[1], (CCV_GET_DATA_TYPE(params.datatype) | params.dim[2]));
 	} else // This won't be recognized by ccv_dense_matrix_t
-		tensor.type = CCV_NO_DATA_ALLOC | CCV_UNMANAGED | CCV_MATRIX_DENSE | params.datatype;
+		tensor.type = CCV_NO_DATA_ALLOC | CCV_UNMANAGED | CCV_MATRIX_DENSE | CCV_GET_DATA_TYPE(params.datatype);
 	tensor.data.u8 = (uint8_t*)ptr;
 	tensor.data_size = 0;
 	return tensor;