diff --git a/backend/app/enhance.py b/backend/app/enhance.py
index 921051d..6bd9737 100644
--- a/backend/app/enhance.py
+++ b/backend/app/enhance.py
@@ -12,7 +12,7 @@ import re
 
 import httpx
 
-from .ollama_client import ollama
+from .ollama_client import OllamaError, ollama
 
 logger = logging.getLogger(__name__)
 
@@ -61,19 +61,29 @@ async def _ask(
     principal : un appel avec des options divergentes force Ollama à recharger
     le modèle en plein message.
     """
-    text = ""
-    async for chunk in ollama.chat(
-        model,
-        [{"role": "system", "content": system},
-         {"role": "user", "content": user}],
-        options={**(options or {}), "temperature": 0.2, "num_predict": num_predict},
-        keep_alive=keep_alive,
-        stream=True,
-    ):
-        text += chunk.get("message", {}).get("content", "")
-        if chunk.get("done"):
-            break
-    return text.strip()
+    messages = [{"role": "system", "content": system},
+                {"role": "user", "content": user}]
+    opts = {**(options or {}), "temperature": 0.2, "num_predict": num_predict}
+    # think=False : un appel utilitaire (plan, critique) ne doit jamais
+    # « réfléchir » — sur un modèle thinking, la pensée dévore le budget et
+    # multiplie la latence. Repli sans le paramètre si le modèle le refuse.
+    think: bool | None = False
+    while True:
+        text = ""
+        try:
+            async for chunk in ollama.chat(
+                model, messages, options=opts, think=think,
+                keep_alive=keep_alive, stream=True,
+            ):
+                text += chunk.get("message", {}).get("content", "")
+                if chunk.get("done"):
+                    break
+            return text.strip()
+        except OllamaError as exc:
+            if think is False and "think" in str(exc).lower():
+                think = None
+                continue
+            raise
 
 
 async def make_plan(
diff --git a/backend/app/memory.py b/backend/app/memory.py
index a6d3ea0..ae01e96 100644
--- a/backend/app/memory.py
+++ b/backend/app/memory.py
@@ -14,7 +14,7 @@ import logging
 import httpx
 
 from . import db
-from .ollama_client import ollama
+from .ollama_client import OllamaError, ollama
 
 logger = logging.getLogger(__name__)
 
@@ -79,20 +79,30 @@ async def maybe_summarize(
         if previous:
             transcript = f"[résumé existant] {previous}\n{transcript}"
 
-        text = ""
-        async for chunk in ollama.chat(
-            model,
-            [
-                {"role": "system", "content": _SUMMARY_PROMPT},
-                {"role": "user", "content": transcript},
-            ],
-            options={**(options or {}), "temperature": 0.2, "num_predict": 350},
-            keep_alive=keep_alive,
-            stream=True,
-        ):
-            text += chunk.get("message", {}).get("content", "")
-            if chunk.get("done"):
+        messages = [
+            {"role": "system", "content": _SUMMARY_PROMPT},
+            {"role": "user", "content": transcript},
+        ]
+        opts = {**(options or {}), "temperature": 0.2, "num_predict": 350}
+        # think=False : le résumé d'arrière-plan ne doit pas « réfléchir »
+        # (latence ×3 sur un modèle thinking). Repli si paramètre refusé.
+        think: bool | None = False
+        while True:
+            text = ""
+            try:
+                async for chunk in ollama.chat(
+                    model, messages, options=opts, think=think,
+                    keep_alive=keep_alive, stream=True,
+                ):
+                    text += chunk.get("message", {}).get("content", "")
+                    if chunk.get("done"):
+                        break
                 break
+            except OllamaError as exc:
+                if think is False and "think" in str(exc).lower():
+                    think = None
+                    continue
+                raise
 
         if text.strip():
             db.set_session_summary(sid, text.strip())
diff --git a/backend/app/routes/chat.py b/backend/app/routes/chat.py
index b61589f..c5a834f 100644
--- a/backend/app/routes/chat.py
+++ b/backend/app/routes/chat.py
@@ -324,7 +324,10 @@ async def chat(req: ChatRequest) -> StreamingResponse:
 
         from ..mcp_client import manager as mcp_manager
 
-        memories, plan, code_model, mcp_tools = await asyncio.gather(
+        # La préparation peut déclencher un CHARGEMENT de modèle (plan, embed)
+        # qui dure plusieurs minutes sur un gros modèle. Sans battement de
+        # cœur, le silence SSE fait couper la connexion par le reverse proxy.
+        prep = asyncio.ensure_future(asyncio.gather(
             rag.recall(req.session_id, req.content, embed_model=cfg.get("embed_model"))
             if want_rag else asyncio.sleep(0, result=[]),
             enhance.make_plan(model, req.content, options=run_opts, keep_alive=keep)
@@ -332,7 +335,15 @@ async def chat(req: ChatRequest) -> StreamingResponse:
             coder.pick_code_model(model, cfg.get("code_model"))
             if use_code else asyncio.sleep(0, result=model),
             mcp_manager.tool_definitions(),
-        )
+        ))
+        while True:
+            try:
+                memories, plan, code_model, mcp_tools = await asyncio.wait_for(
+                    asyncio.shield(prep), timeout=10.0
+                )
+                break
+            except TimeoutError:
+                yield _sse("ping", {"status": "waiting"})
 
         # Pannes MCP éventuelles : notice non bloquante dans le fil.
         for mcp_notice in mcp_manager.notices():