diff --git a/CHANGELOG.md b/CHANGELOG.md
index 5c34e2729..d0a340ecd 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -7,6 +7,8 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+- fix(server): preserve delimiters in string metadata overrides by @allenflux in #2380
+
 ## [0.3.36]
 
 - feat: update llama.cpp to ggml-org/llama.cpp@0c1e57098
diff --git a/llama_cpp/server/model.py b/llama_cpp/server/model.py
index 8aa929202..ca03143fe 100644
--- a/llama_cpp/server/model.py
+++ b/llama_cpp/server/model.py
@@ -236,9 +236,9 @@ def load_llama_from_model_settings(settings: ModelSettings) -> llama_cpp.Llama:
             assert isinstance(settings.kv_overrides, list)
             kv_overrides = {}
             for kv in settings.kv_overrides:
-                key, value = kv.split("=")
+                key, value = kv.split("=", 1)
                 if ":" in value:
-                    value_type, value = value.split(":")
+                    value_type, value = value.split(":", 1)
                     if value_type == "bool":
                         kv_overrides[key] = value.lower() in ["true", "1"]
                     elif value_type == "int":
diff --git a/tests/test_server_model.py b/tests/test_server_model.py
new file mode 100644
index 000000000..daa0f51c1
--- /dev/null
+++ b/tests/test_server_model.py
@@ -0,0 +1,58 @@
+from unittest.mock import Mock
+
+import pytest
+
+import llama_cpp
+from llama_cpp.server.model import LlamaProxy
+from llama_cpp.server.settings import ModelSettings
+
+
+@pytest.mark.parametrize(
+    "value",
+    [
+        "user: ",
+        "{{ messages[0]['role'] == 'user' }}",
+        "user: name=value",
+        "",
+        ":",
+        "=",
+        "用户: name=value",
+    ],
+)
+def test_server_kv_overrides_preserve_string_values(monkeypatch, value):
+    create_model = Mock()
+    monkeypatch.setattr(llama_cpp, "Llama", create_model)
+    settings = ModelSettings(
+        model="unused.gguf",
+        kv_overrides=[f"tokenizer.chat_template=str:{value}"],
+    )
+
+    result = LlamaProxy.load_llama_from_model_settings(settings)
+
+    assert result is create_model.return_value
+    assert create_model.call_args.kwargs["kv_overrides"] == {
+        "tokenizer.chat_template": value
+    }
+
+
+def test_server_kv_overrides_preserve_numeric_and_bool_values(monkeypatch):
+    create_model = Mock()
+    monkeypatch.setattr(llama_cpp, "Llama", create_model)
+    settings = ModelSettings(
+        model="unused.gguf",
+        kv_overrides=[
+            "tokenizer.ggml.add_bos_token=bool:true",
+            "tokenizer.ggml.add_eos_token=bool:false",
+            "llama.context_length=int:2048",
+            "llama.rope.freq_base=float:10000.5",
+        ],
+    )
+
+    LlamaProxy.load_llama_from_model_settings(settings)
+
+    assert create_model.call_args.kwargs["kv_overrides"] == {
+        "tokenizer.ggml.add_bos_token": True,
+        "tokenizer.ggml.add_eos_token": False,
+        "llama.context_length": 2048,
+        "llama.rope.freq_base": 10000.5,
+    }