diff --git a/src/lib/ai/hybridRouter.ts b/src/lib/ai/hybridRouter.ts
index aeea31c..22ba3b1 100644
--- a/src/lib/ai/hybridRouter.ts
+++ b/src/lib/ai/hybridRouter.ts
@@ -65,8 +65,11 @@ async function callOllama(
   endpoint: string,
   options: HybridOptions
 ): Promise<string> {
-  const requestTimeout = options.isBackground ? 120_000 : 60_000;
-  const connectionTimeout = 8_000;
+  const requestTimeout = options.isBackground ? 180_000 : 60_000;
+  // 30s connection timeout (was 8s) — large models like llama3.1:70b
+  // may need to swap into VRAM on AI_001/AI_002, which can exceed 10s
+  // when another model is currently loaded.
+  const connectionTimeout = 30_000;
 
   // Build Ollama OpenAI-compatible payload
   const payload = {