diff --git a/src/transformers/models/qwen3_omni_moe/configuration_qwen3_omni_moe.py b/src/transformers/models/qwen3_omni_moe/configuration_qwen3_omni_moe.py
index 43e6250c0fd4..d434f91fa4f5 100644
--- a/src/transformers/models/qwen3_omni_moe/configuration_qwen3_omni_moe.py
+++ b/src/transformers/models/qwen3_omni_moe/configuration_qwen3_omni_moe.py
@@ -872,6 +872,7 @@ def __init__(
         position_id_per_seconds=25,
         audio_start_token_id=151669,
         speaker_id=None,
+        initializer_range=0.02,
         **kwargs,
     ):
         if code_predictor_config is None:
@@ -907,6 +908,7 @@ def __init__(
         self.audio_start_token_id = audio_start_token_id
         self.vision_start_token_id = vision_start_token_id
         self.speaker_id = speaker_id
+        self.initializer_range = initializer_range
         super().__init__(**kwargs)
 
 
@@ -997,6 +999,7 @@ def __init__(
         upsampling_ratios=(2, 2),
         decoder_dim=1536,
         attention_dropout=0.0,
+        initializer_range=0.02,
         **kwargs,
     ):
         self.codebook_size = codebook_size
@@ -1017,6 +1020,7 @@ def __init__(
         self.decoder_dim = decoder_dim
         self.attention_dropout = attention_dropout
         self.rope_parameters = rope_parameters
+        self.initializer_range = initializer_range
 
         super().__init__(**kwargs)
 
@@ -1087,6 +1091,7 @@ def __init__(
         system_token_id=8948,
         user_token_id=872,
         assistant_token_id=77091,
+        initializer_range=0.02,
         **kwargs,
     ):
         if thinker_config is None:
@@ -1113,6 +1118,7 @@ def __init__(
         self.system_token_id = system_token_id
         self.user_token_id = user_token_id
         self.assistant_token_id = assistant_token_id
+        self.initializer_range = initializer_range
         super().__init__(**kwargs)
 
     def get_text_config(self, decoder=False) -> "PreTrainedConfig":
diff --git a/src/transformers/models/qwen3_omni_moe/modeling_qwen3_omni_moe.py b/src/transformers/models/qwen3_omni_moe/modeling_qwen3_omni_moe.py
index ff3a5ef41c77..21d8c6778df8 100644
--- a/src/transformers/models/qwen3_omni_moe/modeling_qwen3_omni_moe.py
+++ b/src/transformers/models/qwen3_omni_moe/modeling_qwen3_omni_moe.py
@@ -3042,7 +3042,7 @@ def get_input_embeddings(self):
 
 @auto_docstring
 class Qwen3OmniMoeTalkerForConditionalGeneration(Qwen3OmniMoeThinkerTextPreTrainedModel, GenerationMixin):
-    _tied_weights_keys = {"lm_head.weight": "model.embed_tokens.weight"}
+    _tied_weights_keys = {}
     _tp_plan = {"lm_head": "colwise_rep"}
     _pp_plan = {"lm_head": (["hidden_states"], ["logits"])}
     config_class = Qwen3OmniMoeTalkerConfig
diff --git a/src/transformers/models/qwen3_omni_moe/modular_qwen3_omni_moe.py b/src/transformers/models/qwen3_omni_moe/modular_qwen3_omni_moe.py
index 79ae9f95c374..593387111443 100644
--- a/src/transformers/models/qwen3_omni_moe/modular_qwen3_omni_moe.py
+++ b/src/transformers/models/qwen3_omni_moe/modular_qwen3_omni_moe.py
@@ -638,6 +638,7 @@ def __init__(
         position_id_per_seconds=25,
         audio_start_token_id=151669,
         speaker_id=None,
+        initializer_range=0.02,
         **kwargs,
     ):
         if code_predictor_config is None:
@@ -673,6 +674,7 @@ def __init__(
         self.audio_start_token_id = audio_start_token_id
         self.vision_start_token_id = vision_start_token_id
         self.speaker_id = speaker_id
+        self.initializer_range = initializer_range
         super().__init__(**kwargs)
 
 
@@ -763,6 +765,7 @@ def __init__(
         upsampling_ratios=(2, 2),
         decoder_dim=1536,
         attention_dropout=0.0,
+        initializer_range=0.02,
         **kwargs,
     ):
         self.codebook_size = codebook_size
@@ -783,6 +786,7 @@ def __init__(
         self.decoder_dim = decoder_dim
         self.attention_dropout = attention_dropout
         self.rope_parameters = rope_parameters
+        self.initializer_range = initializer_range
 
         super().__init__(**kwargs)
 
@@ -853,6 +857,7 @@ def __init__(
         system_token_id=8948,
         user_token_id=872,
         assistant_token_id=77091,
+        initializer_range=0.02,
         **kwargs,
     ):
         if thinker_config is None:
@@ -879,6 +884,7 @@ def __init__(
         self.system_token_id = system_token_id
         self.user_token_id = user_token_id
         self.assistant_token_id = assistant_token_id
+        self.initializer_range = initializer_range
         super().__init__(**kwargs)
 
     def get_text_config(self, decoder=False) -> "PreTrainedConfig":
@@ -1869,6 +1875,7 @@ def get_input_embeddings(self):
 
 
 class Qwen3OmniMoeTalkerForConditionalGeneration(Qwen3MoeForCausalLM):
+    _tied_weights_keys = {}
     config_class = Qwen3OmniMoeTalkerConfig
     base_model_prefix = "talker"
     _no_split_modules = ["Qwen3OmniMoeTalkerCodePredictorModelForConditionalGeneration"]
diff --git a/tests/models/qwen3_omni_moe/test_configuration_and_loading.py b/tests/models/qwen3_omni_moe/test_configuration_and_loading.py
new file mode 100644
index 000000000000..0f90df2c7a36
--- /dev/null
+++ b/tests/models/qwen3_omni_moe/test_configuration_and_loading.py
@@ -0,0 +1,24 @@
+from transformers.models.qwen3_omni_moe.configuration_qwen3_omni_moe import (
+    Qwen3OmniMoeCode2WavConfig,
+    Qwen3OmniMoeConfig,
+    Qwen3OmniMoeTalkerConfig,
+)
+from transformers.models.qwen3_omni_moe.modeling_qwen3_omni_moe import (
+    Qwen3OmniMoeTalkerForConditionalGeneration,
+)
+
+
+def test_qwen3_omni_moe_configs_have_initializer_range():
+    talker_config = Qwen3OmniMoeTalkerConfig()
+    assert hasattr(talker_config, "initializer_range")
+
+    code2wav_config = Qwen3OmniMoeCode2WavConfig()
+    assert hasattr(code2wav_config, "initializer_range")
+
+    main_config = Qwen3OmniMoeConfig()
+    assert hasattr(main_config, "initializer_range")
+
+
+def test_qwen3_omni_moe_talker_has_no_tied_weights():
+    tied_keys = Qwen3OmniMoeTalkerForConditionalGeneration._tied_weights_keys
+    assert tied_keys in (None, {})