diff --git a/convert_torch2paddle.py b/convert_torch2paddle.py
new file mode 100644
index 000000000..b13a793a8
--- /dev/null
+++ b/convert_torch2paddle.py
@@ -0,0 +1,18 @@
+import paddle
+import torch
+
+torch_model_dict = torch.load('large-v3-turbo.pt')['model_state_dict']
+
+paddle_model_state_dict = {}
+for key, val in torch_model_dict.items():
+    if key.endswith(
+            'weight'
+    ) and val.ndim == 2 and key != "decoder.token_embedding.weight":
+        val = val.T
+    paddle_model_state_dict[key] = paddle.to_tensor(
+        val.cpu().numpy()).astype("float32")
+
+# add other params in case if need, such as:
+paddle_model_state_dict['dims'] = torch.load('large-v3-turbo.pt')['dims']
+
+paddle.save(paddle_model_state_dict, 'weights.params')