diff --git a/backend/shonar/services/ai/ollama.py b/backend/shonar/services/ai/ollama.py
index 574f77a..dd293a5 100644
--- a/backend/shonar/services/ai/ollama.py
+++ b/backend/shonar/services/ai/ollama.py
@@ -18,7 +18,7 @@ class OllamaProvider:
         self,
         base_url: str,
         model: str = "",
-        timeout_s: float = 300.0,
+        timeout_s: float = 900.0,
         http_client: httpx.AsyncClient | None = None,
     ) -> None:
         if not base_url.strip():
@@ -43,6 +43,12 @@ class OllamaProvider:
             "model": self.model,
             "stream": False,
             "format": "json",
+            # qwen3-family models "think" by default: a long reasoning
+            # chain before the JSON answer, brutally slow on CPU and it
+            # does not improve the summary. Ask for the answer directly
+            # (ignored by non-thinking models).
+            "think": False,
+            "options": {"num_ctx": 8192},
             "messages": [
                 {"role": "system", "content": SYSTEM_PROMPT},
                 {"role": "user", "content": build_user_message(transcript, title)},