diff --git a/fusion_accounting_bank_rec/tests/__init__.py b/fusion_accounting_bank_rec/tests/__init__.py
index 63bfa701..d98aea8f 100644
--- a/fusion_accounting_bank_rec/tests/__init__.py
+++ b/fusion_accounting_bank_rec/tests/__init__.py
@@ -22,3 +22,4 @@ from . import test_migration_round_trip
 from . import test_coexistence
 from . import test_bank_rec_tours
 from . import test_performance_benchmarks
+from . import test_local_llm_compat
diff --git a/fusion_accounting_bank_rec/tests/test_local_llm_compat.py b/fusion_accounting_bank_rec/tests/test_local_llm_compat.py
new file mode 100644
index 00000000..97e94293
--- /dev/null
+++ b/fusion_accounting_bank_rec/tests/test_local_llm_compat.py
@@ -0,0 +1,102 @@
+"""Local LLM compatibility test (LM Studio, Ollama, etc.).
+
+Skips if no local OpenAI-compatible LLM server is reachable. When one is
+running (LM Studio at :1234, Ollama at :11434), runs an end-to-end:
+
+1. Configure ``ir.config_parameter`` to point at the local server.
+2. Trigger ``engine.suggest_matches`` with the 'openai' provider.
+3. Assert the call did not crash and produced at least one suggestion.
+
+The smoke is intentionally lenient: local models often emit malformed
+JSON, in which case ``confidence_scoring`` falls back to statistical-only
+ranking. We assert end-to-end happiness, not AI re-rank quality.
+"""
+
+import socket
+
+from odoo.tests.common import TransactionCase, tagged
+
+from . import _factories as f
+
+
+def _server_reachable(host, port, timeout=1.0):
+    try:
+        with socket.create_connection((host, port), timeout=timeout):
+            return True
+    except (OSError, socket.timeout):
+        return False
+
+
+def _detect_local_llm():
+    """Return (base_url, model_name) tuple, or (None, None) if no server.
+
+    Tries LM Studio (:1234) and Ollama (:11434) on both
+    ``host.docker.internal`` (so the container can reach the host) and
+    ``localhost`` (so a non-containerised run finds the same servers).
+    """
+    candidates = (
+        ('host.docker.internal', 1234, 'local-model'),    # LM Studio
+        ('host.docker.internal', 11434, 'llama3.1:8b'),   # Ollama
+        ('localhost', 1234, 'local-model'),
+        ('localhost', 11434, 'llama3.1:8b'),
+    )
+    for host, port, default_model in candidates:
+        if _server_reachable(host, port, timeout=0.5):
+            return (f'http://{host}:{port}/v1', default_model)
+    return (None, None)
+
+
+@tagged('post_install', '-at_install', 'local_llm')
+class TestLocalLLMCompat(TransactionCase):
+
+    def setUp(self):
+        super().setUp()
+        self.base_url, self.model = _detect_local_llm()
+        if not self.base_url:
+            self.skipTest(
+                "No local LLM server detected "
+                "(LM Studio :1234 / Ollama :11434)")
+
+    def test_suggest_matches_with_local_llm(self):
+        params = self.env['ir.config_parameter'].sudo()
+        prior = {
+            'fusion_accounting.openai_base_url': params.get_param(
+                'fusion_accounting.openai_base_url'),
+            'fusion_accounting.openai_model': params.get_param(
+                'fusion_accounting.openai_model'),
+            'fusion_accounting.openai_api_key': params.get_param(
+                'fusion_accounting.openai_api_key'),
+            'fusion_accounting.provider.bank_rec_suggest': params.get_param(
+                'fusion_accounting.provider.bank_rec_suggest'),
+        }
+
+        params.set_param('fusion_accounting.openai_base_url', self.base_url)
+        params.set_param('fusion_accounting.openai_model', self.model)
+        # Local servers ignore the key but the adapter requires *some* value.
+        params.set_param('fusion_accounting.openai_api_key', 'lm-studio')
+        params.set_param(
+            'fusion_accounting.provider.bank_rec_suggest', 'openai')
+
+        try:
+            partner = self.env['res.partner'].create(
+                {'name': 'Local LLM Partner'})
+            f.make_invoice(self.env, partner=partner, amount=750)
+            bank_line = f.make_bank_line(
+                self.env, amount=750, partner=partner,
+                memo='REF 12345 Local LLM test')
+
+            result = self.env['fusion.reconcile.engine'].suggest_matches(
+                bank_line, limit_per_line=3)
+
+            self.assertIn(bank_line.id, result)
+            suggestions = self.env['fusion.reconcile.suggestion'].search([
+                ('statement_line_id', '=', bank_line.id),
+            ])
+            self.assertGreater(
+                len(suggestions), 0,
+                "Local LLM run should still produce at least one suggestion "
+                "(statistical fallback if AI re-rank fails)")
+        finally:
+            for key, value in prior.items():
+                if value is not None:
+                    params.set_param(key, value)