diff --git a/src/gluonnlp/models/bart.py b/src/gluonnlp/models/bart.py
index 3f2ae159c9..b6f7a7c32d 100644
--- a/src/gluonnlp/models/bart.py
+++ b/src/gluonnlp/models/bart.py
@@ -315,13 +315,19 @@ def get_pretrained_bart(model_name: str = 'fairseq_roberta_base',
     assert model_name in PRETRAINED_URL, '{} is not found. All available are {}'.format(
         model_name, list_pretrained_bart())
     cfg_path = PRETRAINED_URL[model_name]['cfg']
+    if isinstance(cfg_path, CN):
+        cfg = cfg_path
+    else:
+        cfg = None
     merges_path = PRETRAINED_URL[model_name]['merges']
     vocab_path = PRETRAINED_URL[model_name]['vocab']
     params_path = PRETRAINED_URL[model_name]['params']
 
     local_paths = dict()
-    for k, path in [('cfg', cfg_path), ('vocab', vocab_path),
-                    ('merges', merges_path)]:
+    download_jobs = [('vocab', vocab_path), ('merges', merges_path)]
+    if cfg is None:
+        download_jobs.append(('cfg', cfg_path))
+    for k, path in download_jobs:
         local_paths[k] = download(url=get_repo_model_zoo_url() + path,
                                   path=os.path.join(root, path),
                                   sha1_hash=FILE_STATS[path])
@@ -339,7 +345,8 @@ def get_pretrained_bart(model_name: str = 'fairseq_roberta_base',
         merges_file=local_paths['merges'],
         vocab_file=local_paths['vocab'],
         lowercase=do_lower)
-    cfg = BartModel.get_cfg().clone_merge(local_paths['cfg'])
+    if cfg is None:
+        cfg = BartModel.get_cfg().clone_merge(local_paths['cfg'])
     return cfg, tokenizer, local_params_path, local_mlm_params_path