diff --git a/mock/app.js b/mock/app.js
index a2c07e0..1bcdacb 100644
--- a/mock/app.js
+++ b/mock/app.js
@@ -122,7 +122,11 @@ async function warmLLM() {
   } catch {}
   warming = false;
 }
-setInterval(() => { if (llmOn && !scriptBusy) warmLLM(); }, 120000);   // keep the model warm while the tab is open
+// Keep-alive is DISABLED by default: the probe pins the big model (Qwen3.8-27B-dual)
+// on the LM Studio server and prevents loading other models. Opt in per session
+// with ?warm=1 in the URL (accepts the cold first turn cost, see above).
+const warmEnabled = new URLSearchParams(location.search).has('warm');
+setInterval(() => { if (warmEnabled && llmOn && !scriptBusy) warmLLM(); }, 120000);
 // ---------------- web search client (SearXNG via the /search proxy) ------
 // Live facts (opening hours, fees, seasonal notes) with provenance URLs —
 // the third leg of the 3-source blend. Enrichment APPENDS to suggestion