diff --git a/hosts/galadriel/configuration.nix b/hosts/galadriel/configuration.nix
index 73631f4..9dfe296 100644
--- a/hosts/galadriel/configuration.nix
+++ b/hosts/galadriel/configuration.nix
@@ -36,8 +36,10 @@
     ../../modules/qbittorrent.nix
     ../../modules/mealie.nix
     ../../modules/miniflux.nix
-    ../../modules/ollama.nix
-    ../../modules/openwebui.nix
+    #../../modules/ollama.nix     # replaced by llama-cpp + llama-swap
+    #../../modules/openwebui.nix  # using llama-cpp built-in UI instead
+    ../../modules/llama-cpp.nix
+    ../../modules/llama-swap.nix
     ../../modules/librechat.nix
     ../../modules/immich.nix
 
diff --git a/modules/llama-cpp.nix b/modules/llama-cpp.nix
new file mode 100644
index 0000000..0ae2ddd
--- /dev/null
+++ b/modules/llama-cpp.nix
@@ -0,0 +1,91 @@
+{
+  config,
+  pkgs,
+  lib,
+  ...
+}:
+
+let
+  modelDir = "/var/lib/llama-cpp/models";
+
+  # llama-cpp with Vulkan support for Intel Arc
+  llama-cpp-vulkan = pkgs.llama-cpp.override {
+    vulkanSupport = true;
+    cudaSupport = false;
+    rocmSupport = false;
+    openclSupport = false;
+    blasSupport = true;
+  };
+
+  llama-server = lib.getExe' llama-cpp-vulkan "llama-server";
+
+  # Model definitions: name -> { url, filename }
+  models = {
+    "Ministral-3-8B-Reasoning-Q4_K_M" = {
+      url = "https://huggingface.co/mistralai/Ministral-3-8B-Reasoning-2512-GGUF/resolve/main/Ministral-3-8B-Reasoning-2512-Q4_K_M.gguf";
+      filename = "Ministral-3-8B-Reasoning-2512-Q4_K_M.gguf";
+    };
+    "Qwen3.5-35B-A3B-UD-Q3_K_M" = {
+      url = "https://huggingface.co/unsloth/Qwen3.5-35B-A3B-GGUF/resolve/main/Qwen3.5-35B-A3B-UD-Q3_K_M.gguf";
+      filename = "Qwen3.5-35B-A3B-UD-Q3_K_M.gguf";
+    };
+    "Qwen3.5-27B-UD-Q4_K_XL" = {
+      url = "https://huggingface.co/unsloth/Qwen3.5-27B-GGUF/resolve/main/Qwen3.5-27B-UD-Q4_K_XL.gguf";
+      filename = "Qwen3.5-27B-UD-Q4_K_XL.gguf";
+    };
+    "LFM2-24B-A2B-Q4_K_M" = {
+      url = "https://huggingface.co/LiquidAI/LFM2-24B-A2B-GGUF/resolve/main/LFM2-24B-A2B-Q4_K_M.gguf";
+      filename = "LFM2-24B-A2B-Q4_K_M.gguf";
+    };
+    "Nanbeige4-3B-Thinking-Q4_K_M" = {
+      url = "https://huggingface.co/bartowski/Nanbeige_Nanbeige4-3B-Thinking-2511-GGUF/resolve/main/Nanbeige_Nanbeige4-3B-Thinking-2511-Q4_K_M.gguf";
+      filename = "Nanbeige_Nanbeige4-3B-Thinking-2511-Q4_K_M.gguf";
+    };
+  };
+
+  # Generate a systemd oneshot service per model that downloads if missing
+  downloadServices = lib.mapAttrs' (
+    name: model:
+    lib.nameValuePair "llama-cpp-download-${name}" {
+      description = "Download GGUF model: ${name}";
+      after = [ "network-online.target" ];
+      wants = [ "network-online.target" ];
+      wantedBy = [ "multi-user.target" ];
+
+      serviceConfig = {
+        Type = "oneshot";
+        RemainAfterExit = true;
+        User = "llama-cpp";
+        Group = "llama-cpp";
+        ExecStart = pkgs.writeShellScript "download-${name}" ''
+          MODEL_PATH="${modelDir}/${model.filename}"
+          if [ ! -f "$MODEL_PATH" ]; then
+            echo "Downloading ${name}..."
+            ${lib.getExe pkgs.curl} -L -o "$MODEL_PATH.tmp" "${model.url}"
+            mv "$MODEL_PATH.tmp" "$MODEL_PATH"
+            echo "Download complete: ${name}"
+          else
+            echo "Model already exists: ${name}"
+          fi
+        '';
+      };
+    }
+  ) models;
+in
+{
+  environment.systemPackages = [ llama-cpp-vulkan ];
+
+  users.users.llama-cpp = {
+    isSystemUser = true;
+    group = "llama-cpp";
+    home = "/var/lib/llama-cpp";
+    description = "llama-cpp service user";
+  };
+  users.groups.llama-cpp = { };
+
+  systemd.tmpfiles.rules = [
+    "d ${modelDir} 0755 llama-cpp llama-cpp - -"
+  ];
+
+  systemd.services = downloadServices;
+}
diff --git a/modules/llama-swap.nix b/modules/llama-swap.nix
new file mode 100644
index 0000000..f995ac1
--- /dev/null
+++ b/modules/llama-swap.nix
@@ -0,0 +1,74 @@
+{
+  config,
+  pkgs,
+  lib,
+  ...
+}:
+
+let
+  modelDir = "/var/lib/llama-cpp/models";
+
+  # llama-cpp with Vulkan for Intel Arc
+  llama-cpp-vulkan = pkgs.llama-cpp.override {
+    vulkanSupport = true;
+    cudaSupport = false;
+    rocmSupport = false;
+    openclSupport = false;
+    blasSupport = true;
+  };
+
+  llama-server = lib.getExe' llama-cpp-vulkan "llama-server";
+
+  # Common flags for all models
+  # --no-webui is NOT set so the llama-cpp UI is available per-model
+  commonFlags = port: "-ngl 99 --port \${PORT} --host 127.0.0.1";
+in
+{
+  services.llama-swap = {
+    enable = true;
+    port = 11111;
+    openFirewall = true;
+
+    settings = {
+      healthCheckTimeout = 120;
+
+      models = {
+        "ministral-3-8b-reasoning" = {
+          cmd = "${llama-server} --port \${PORT} --host 127.0.0.1 -ngl 99 -m ${modelDir}/Ministral-3-8B-Reasoning-2512-Q4_K_M.gguf -c 32768";
+          aliases = [
+            "ministral-3"
+            "ministral"
+          ];
+        };
+        "qwen3.5-35b-a3b" = {
+          cmd = "${llama-server} --port \${PORT} --host 127.0.0.1 -ngl 99 -m ${modelDir}/Qwen3.5-35B-A3B-UD-Q3_K_M.gguf -c 32768";
+          aliases = [
+            "qwen3.5-35b"
+            "qwen-moe"
+          ];
+        };
+        "qwen3.5-27b" = {
+          cmd = "${llama-server} --port \${PORT} --host 127.0.0.1 -ngl 99 -m ${modelDir}/Qwen3.5-27B-UD-Q4_K_XL.gguf -c 16384";
+          aliases = [
+            "qwen3.5"
+            "qwen-27b"
+          ];
+        };
+        "lfm2-24b-a2b" = {
+          cmd = "${llama-server} --port \${PORT} --host 127.0.0.1 -ngl 99 -m ${modelDir}/LFM2-24B-A2B-Q4_K_M.gguf -c 32768";
+          aliases = [
+            "lfm2"
+            "liquid"
+          ];
+        };
+        "nanbeige4-3b-thinking" = {
+          cmd = "${llama-server} --port \${PORT} --host 127.0.0.1 -ngl 99 -m ${modelDir}/Nanbeige_Nanbeige4-3B-Thinking-2511-Q4_K_M.gguf -c 32768";
+          aliases = [
+            "nanbeige"
+            "nanbeige4"
+          ];
+        };
+      };
+    };
+  };
+}