diff --git a/litellm/main.py b/litellm/main.py
index 73db4a11cb7..3a338799c1f 100644
--- a/litellm/main.py
+++ b/litellm/main.py
@@ -3969,8 +3969,8 @@ def completion(  # type: ignore # noqa: PLR0915
             response = model_response
         elif custom_llm_provider == "ollama":
             api_base = (
-                litellm.api_base
-                or api_base
+                api_base
+                or litellm.api_base
                 or get_secret("OLLAMA_API_BASE")
                 or "http://localhost:11434"
             )
@@ -3998,8 +3998,8 @@ def completion(  # type: ignore # noqa: PLR0915
 
         elif custom_llm_provider == "ollama_chat":
             api_base = (
-                litellm.api_base
-                or api_base
+                api_base
+                or litellm.api_base
                 or get_secret("OLLAMA_API_BASE")
                 or "http://localhost:11434"
             )
@@ -5317,8 +5317,8 @@ def embedding(  # noqa: PLR0915
             )
         elif custom_llm_provider == "ollama":
             api_base = (
-                litellm.api_base
-                or api_base
+                api_base
+                or litellm.api_base
                 or get_secret_str("OLLAMA_API_BASE")
                 or "http://localhost:11434"
             )  # type: ignore
diff --git a/tests/local_testing/test_completion.py b/tests/local_testing/test_completion.py
index d6370bfd049..b4d6e205a76 100644
--- a/tests/local_testing/test_completion.py
+++ b/tests/local_testing/test_completion.py
@@ -1323,6 +1323,37 @@ def test_lm_studio_completion(monkeypatch):
         print(e)
 
 
+@pytest.mark.parametrize("model", ["ollama/phi", "ollama_chat/phi"])
+def test_ollama_completion_explicit_api_base_takes_precedence(monkeypatch, model):
+    monkeypatch.setattr(litellm, "api_base", "https://api.deepseek.com")
+
+    with patch("litellm.main.base_llm_http_handler.completion") as mock_completion:
+        mock_completion.return_value = MagicMock()
+
+        litellm.completion(
+            model=model,
+            messages=[{"role": "user", "content": "Hello"}],
+            api_base="http://localhost:11434",
+        )
+
+        assert mock_completion.call_args.kwargs["api_base"] == "http://localhost:11434"
+
+
+def test_ollama_embedding_explicit_api_base_takes_precedence(monkeypatch):
+    monkeypatch.setattr(litellm, "api_base", "https://api.deepseek.com")
+
+    with patch("litellm.main.ollama.ollama_embeddings") as mock_embeddings:
+        mock_embeddings.return_value = MagicMock()
+
+        litellm.embedding(
+            model="ollama/qwen3-embedding:0.6b",
+            input="hello",
+            api_base="http://localhost:11434",
+        )
+
+        assert mock_embeddings.call_args.kwargs["api_base"] == "http://localhost:11434"
+
+
 # ################### Hugging Face Conversational models ########################
 # def hf_test_completion_conv():
 #     try: