diff --git a/llama/compat/README.md b/llama/compat/README.md
index 8c356a7dd..cf5f29dcd 100644
--- a/llama/compat/README.md
+++ b/llama/compat/README.md
@@ -28,6 +28,8 @@ intentionally skipped so a developer can iterate on a local llama.cpp tree.
 - `compat.cmake`, `apply-patch.cmake` - CMake glue and an idempotent applier
   (used by `llama/server/CMakeLists.txt`) that applies every `*.patch` under
   this directory — the hooks patch plus each `models/` architecture patch.
+- `ui/llama-*-ui.tar.gz` - prebuilt llama.cpp server UI assets from the pinned
+  release, extracted into the fetched source before upstream embeds them.
 - `models/` - the sibling **new-architecture** layer: implementations of
   architectures llama.cpp doesn't support yet, each added via a small
   registration patch. (Those files *add* archs; the files above *translate*
diff --git a/llama/compat/apply-patch.cmake b/llama/compat/apply-patch.cmake
index e7ca91af3..e9b607013 100644
--- a/llama/compat/apply-patch.cmake
+++ b/llama/compat/apply-patch.cmake
@@ -48,3 +48,41 @@ foreach(PATCH_FILE IN LISTS _patches)
 
     message(STATUS "llama/compat: applied patch")
 endforeach()
+
+set(_ui_archive "")
+set(_llama_cpp_version_file "${PATCH_DIR}/../../LLAMA_CPP_VERSION")
+if(EXISTS "${_llama_cpp_version_file}")
+    file(READ "${_llama_cpp_version_file}" _llama_cpp_tag)
+    string(STRIP "${_llama_cpp_tag}" _llama_cpp_tag)
+    set(_ui_archive "${PATCH_DIR}/ui/llama-${_llama_cpp_tag}-ui.tar.gz")
+endif()
+
+if(_ui_archive AND EXISTS "${_ui_archive}")
+    set(_ui_dist_dir "${_patch_workdir}/tools/ui/dist")
+
+    if(EXISTS "${_ui_dist_dir}/index.html")
+        message(STATUS "llama/compat: bundled UI assets already present, skipping")
+    else()
+        set(_ui_extract_dir "${_patch_workdir}/tools/ui/.ollama-ui-extract")
+        file(REMOVE_RECURSE "${_ui_extract_dir}")
+        file(MAKE_DIRECTORY "${_ui_dist_dir}" "${_ui_extract_dir}")
+        file(ARCHIVE_EXTRACT INPUT "${_ui_archive}" DESTINATION "${_ui_extract_dir}")
+
+        file(GLOB _ui_roots RELATIVE "${_ui_extract_dir}" "${_ui_extract_dir}/*")
+        list(LENGTH _ui_roots _ui_root_count)
+        if(NOT _ui_root_count EQUAL 1)
+            message(FATAL_ERROR
+                "llama/compat: expected one top-level directory in ${_ui_archive}")
+        endif()
+
+        list(GET _ui_roots 0 _ui_root)
+        if(NOT EXISTS "${_ui_extract_dir}/${_ui_root}/index.html")
+            message(FATAL_ERROR
+                "llama/compat: ${_ui_archive} does not contain UI assets")
+        endif()
+
+        file(COPY "${_ui_extract_dir}/${_ui_root}/" DESTINATION "${_ui_dist_dir}")
+        file(REMOVE_RECURSE "${_ui_extract_dir}")
+        message(STATUS "llama/compat: extracted bundled UI assets from ${_ui_archive}")
+    endif()
+endif()
diff --git a/llama/compat/llama-cpp-ui-empty-assets.patch b/llama/compat/llama-cpp-ui-empty-assets.patch
deleted file mode 100644
index 31e90898f..000000000
--- a/llama/compat/llama-cpp-ui-empty-assets.patch
+++ /dev/null
@@ -1,15 +0,0 @@
-diff --git a/tools/ui/embed.cpp b/tools/ui/embed.cpp
---- a/tools/ui/embed.cpp
-+++ b/tools/ui/embed.cpp
-@@ -147,9 +147,9 @@ int main(int argc, char ** argv) {
- 
-     const std::string out_cpp   = argv[1];
-     const std::string out_h     = argv[2];
--    const std::string asset_dir = argv[3];
-+    const std::string asset_dir = argc == 4 ? argv[3] : "";
- 
--    const bool        use_gzip = std::filesystem::exists(asset_dir + "/_gzip");
-+    const bool        use_gzip = !asset_dir.empty() && std::filesystem::exists(asset_dir + "/_gzip");
-     const std::string in_dir   = use_gzip ? (asset_dir + "/_gzip") : asset_dir;
- 
-     std::vector<asset_entry> assets;
diff --git a/llama/compat/ui/llama-b9626-ui.tar.gz b/llama/compat/ui/llama-b9626-ui.tar.gz
new file mode 100644
index 000000000..50f845ea8
Binary files /dev/null and b/llama/compat/ui/llama-b9626-ui.tar.gz differ