diff --git a/src/llamacppctl/http_ops.py b/src/llamacppctl/http_ops.py
index 8fb2abf..f414914 100644
--- a/src/llamacppctl/http_ops.py
+++ b/src/llamacppctl/http_ops.py
@@ -145,6 +145,10 @@ def stream_chat(
     if resp.status_code != 200:
         raise HttpError(f"chat completion returned HTTP {resp.status_code}: {resp.text[:500]}")
 
+    # llama.cpp streams text/event-stream WITHOUT a charset; requests would then
+    # fall back to a non-UTF-8 decode and mangle multibyte chars (Umlaute).
+    resp.encoding = "utf-8"
+
     finish_reason = ""
     for line in resp.iter_lines(decode_unicode=True):
         if not line or not line.startswith("data:"):
diff --git a/tests/test_http_ops.py b/tests/test_http_ops.py
index 0085d1a..99fac57 100644
--- a/tests/test_http_ops.py
+++ b/tests/test_http_ops.py
@@ -163,6 +163,8 @@ def test_stream_chat_yields_content_and_finish():
     finishes = [v for k, v in events if k == "finish"]
     assert "".join(contents) == "Hallo"
     assert finishes == ["stop"]
+    # Stream must be forced to UTF-8 (llama.cpp sends no charset -> Umlaut-Bug).
+    assert mock_resp.encoding == "utf-8"
 
 
 def test_wait_until_ready_succeeds_immediately():