diff --git a/llama_stack/providers/tests/inference/ci_test.py b/llama_stack/providers/tests/inference/ci_test.py
index c28b0be78..3671cec51 100644
--- a/llama_stack/providers/tests/inference/ci_test.py
+++ b/llama_stack/providers/tests/inference/ci_test.py
@@ -28,7 +28,7 @@ CATEGORY_FUNCTIONALITY_TESTS = {
 }
 
 
-def generate_pytest_args(category, provider, env_key):
+def generate_pytest_args(category, provider, test_keywords, env_key):
     test_path = (
         "./llama_stack/providers/tests/inference/test_{model_type}_inference.py".format(
             model_type=category
@@ -37,11 +37,14 @@ def generate_pytest_args(category, provider, env_key):
     pytest_args = [
         test_path,
         "-v",
-        "-s",
+        # "-s",
         "-k",
-        provider,
+        "{provider} and ({test_keywords})".format(
+            provider=provider, test_keywords=" or ".join(test_keywords)
+        ),
         "--inference-model={model_name}".format(model_name=TEST_MODELS[category]),
     ]
+
     if env_key is not None:
         pytest_args.extend(
             [
@@ -57,27 +60,26 @@ def generate_pytest_args(category, provider, env_key):
 def main():
     test_result = []
 
-    for model_category, functionality_tests in CATEGORY_FUNCTIONALITY_TESTS.items():
+    for model_category, test_keywords in CATEGORY_FUNCTIONALITY_TESTS.items():
         for provider, env_key in INFERENCE_PROVIDER_ENV_KEY.items():
             if provider == "ollama":
-                proc = subprocess.Popen(
-                    [
-                        "ollama",
-                        "run",
-                        (
-                            "llama3.1:8b-instruct-fp16"
-                            if model_category == "text"
-                            else "llama3.2-vision"
-                        ),
-                    ]
+                ollama_model_alias = (
+                    "llama3.1:8b-instruct-fp16"
+                    if model_category == "text"
+                    else "llama3.2-vision:11b-instruct-fp16"
                 )
+                proc = subprocess.Popen(["ollama", "run", ollama_model_alias])
                 retcode = pytest.main(
-                    generate_pytest_args(model_category, provider, env_key)
+                    generate_pytest_args(
+                        model_category, provider, test_keywords, env_key
+                    )
                 )
-                proc.terminate()
+                proc = subprocess.Popen(["ollama", "stop", ollama_model_alias])
             else:
                 retcode = pytest.main(
-                    generate_pytest_args(model_category, provider, env_key)
+                    generate_pytest_args(
+                        model_category, provider, test_keywords, env_key
+                    )
                 )