diff --git a/sqlmesh/utils/jinja.py b/sqlmesh/utils/jinja.py
index c9339cf404..508c6dce2d 100644
--- a/sqlmesh/utils/jinja.py
+++ b/sqlmesh/utils/jinja.py
@@ -206,6 +206,20 @@ def extract_macro_references_and_variables(
     return macro_references, variables
 
 
+def sort_dict_recursive(
+    item: t.Dict[str, t.Any],
+) -> t.Dict[str, t.Any]:
+    sorted_dict: t.Dict[str, t.Any] = {}
+    for k, v in sorted(item.items()):
+        if isinstance(v, list):
+            sorted_dict[k] = sorted(v)
+        elif isinstance(v, dict):
+            sorted_dict[k] = sort_dict_recursive(v)
+        else:
+            sorted_dict[k] = v
+    return sorted_dict
+
+
 JinjaGlobalAttribute = t.Union[str, int, float, bool, AttributeDict]
 
 
@@ -440,7 +454,7 @@ def to_expressions(self) -> t.List[Expression]:
                 d.PythonCode(
                     expressions=[
                         f"{k} = '{v}'" if isinstance(v, str) else f"{k} = {v}"
-                        for k, v in sorted(filtered_objs.items())
+                        for k, v in sort_dict_recursive(filtered_objs).items()
                     ]
                 )
             )
diff --git a/tests/utils/test_jinja.py b/tests/utils/test_jinja.py
index 5eb00aeb3c..1cf7c1bf95 100644
--- a/tests/utils/test_jinja.py
+++ b/tests/utils/test_jinja.py
@@ -302,3 +302,30 @@ def test_dbt_adapter_macro_scope():
 
     rendered = registry.build_environment().from_string("{{ spark__macro_a() }}").render()
     assert rendered.strip() == "macro_a"
+
+
+def test_macro_registry_to_expressions_sorted():
+    refs = AttributeDict(
+        {
+            "payments": {
+                "database": "jaffle_shop",
+                "schema": "main",
+                "nested": {"foo": "bar", "baz": "bing"},
+            },
+            "orders": {"schema": "main", "database": "jaffle_shop", "nested_list": ["b", "a", "c"]},
+        }
+    )
+
+    registry = JinjaMacroRegistry()
+    registry.add_globals({"sources": {}, "refs": refs})
+
+    # Ensure that the AttributeDict string representation is sorted
+    # in order to prevent an unexpected *visual* diff in ModelDiff
+    # (note that the actual diff is based on the data hashes, so this is purely visual)
+    expressions = registry.to_expressions()
+    assert len(expressions) == 1
+    assert (
+        expressions[0].sql(dialect="duckdb")
+        == "refs = {'orders': {'database': 'jaffle_shop', 'nested_list': ['a', 'b', 'c'], 'schema': 'main'}, 'payments': {'database': 'jaffle_shop', 'nested': {'baz': 'bing', 'foo': 'bar'}, 'schema': 'main'}}\n"
+        "sources = {}"
+    )