diff --git a/tests/openvino/test_modeling.py b/tests/openvino/test_modeling.py
index b3b428c6a5..1eabb4487e 100644
--- a/tests/openvino/test_modeling.py
+++ b/tests/openvino/test_modeling.py
@@ -842,7 +842,8 @@ def test_compare_to_transformers(self, model_arch):
             transformers_outputs = transformers_model(**tokens)
 
         # Compare tensor outputs
-        self.assertTrue(torch.allclose(ov_outputs.logits, transformers_outputs.logits, equal_nan=True, atol=1e-4))
+        atol = 1e-3 if model_arch == "minicpm" else 1e-4
+        self.assertTrue(torch.allclose(ov_outputs.logits, transformers_outputs.logits, equal_nan=True, atol=atol))
 
         # Qwen tokenizer does not support padding
 
diff --git a/tests/openvino/test_quantization.py b/tests/openvino/test_quantization.py
index 9f73059e02..73cb8f9611 100644
--- a/tests/openvino/test_quantization.py
+++ b/tests/openvino/test_quantization.py
@@ -745,7 +745,7 @@ def preprocess_function(examples, tokenizer):
 
 
 class OVTrainerTest(unittest.TestCase):
-    SUPPORTED_ARCHITECTURES_WITH_EXPECTED_QUANTIZED_MATMULS = (("albert", 63, 39),)
+    SUPPORTED_ARCHITECTURES_WITH_EXPECTED_QUANTIZED_MATMULS = (("albert", 64, 39),)
 
     @parameterized.expand(SUPPORTED_ARCHITECTURES_WITH_EXPECTED_QUANTIZED_MATMULS)
     def test_aware_training_quantization(self, model_name, expected_fake_quantize, expected_int8):