diff --git a/ansible/host_vars/ci-01.lab.alexpires.me.yml b/ansible/host_vars/ci-01.lab.alexpires.me.yml
index 8d2cad6..1cfcb39 100644
--- a/ansible/host_vars/ci-01.lab.alexpires.me.yml
+++ b/ansible/host_vars/ci-01.lab.alexpires.me.yml
@@ -29,3 +29,58 @@ windows_exporter_collectors:
   - tcp
   - time
   - update
+
+llama_server_windows_install_dir: "D:\\llama_server"
+llama_server_windows_user: "llama"
+llama_server_windows_working_dir: "D:\\llama_server"
+llama_server_cuda_version: "13.3"
+
+llama_server_models:
+  - name: "luau-qwen3-4b-fim-v0.1"
+    model_id: "mradermacher/Luau-Qwen3-4B-FIM-v0.1-GGUF"
+    quant: "IQ4_XS"
+    revision: "main"
+  - name: "ministral-3-3b-instruct"
+    model_id: "unsloth/Ministral-3-3B-Instruct-2512-GGUF"
+    quant: "IQ4_XS"
+    revision: "main"
+  - name: "qwen2.5-coder-1.5b-instruct"
+    model_id: "Qwen/Qwen2.5-Coder-1.5B-Instruct-GGUF"
+    quant: "Q4_K_M"
+    revision: "main"
+  - name: "qwen2.5-coder-3b-instruct"
+    model_id: "unsloth/Qwen2.5-Coder-3B-Instruct-128K-GGUF"
+    quant: "Q4_K_M"
+    revision: "main"
+
+llama_server_host: "0.0.0.0"
+
+llama_server_windows_sources:
+  - 10.19.4.136
+
+llama_server_args_extra:
+  [
+    "-t",
+    8,
+    "-np",
+    1,
+    "-c",
+    8192,
+    "-b",
+    1024,
+    "-ub",
+    512,
+    "-fa",
+    "on",
+    "-ctk",
+    "q4_0",
+    "-ctv",
+    "q4_0",
+    "-cram",
+    0,
+    "-fit",
+    "off",
+    "--mmap",
+  ]
+
+llama_server_windows_backend: "vulkan"
diff --git a/ansible/host_vars/gpu-01.lab.alexpires.me.yml b/ansible/host_vars/gpu-01.lab.alexpires.me.yml
index 1607131..2f929f2 100644
--- a/ansible/host_vars/gpu-01.lab.alexpires.me.yml
+++ b/ansible/host_vars/gpu-01.lab.alexpires.me.yml
@@ -168,6 +168,37 @@ llama_server_models:
     quant: "IQ4_XS"
     revision: "main"
 
+llama_server_args_extra:
+  [
+    "-t",
+    10,
+    "-np",
+    1,
+    "-b",
+    1024,
+    "-ub",
+    512,
+    "-fa",
+    "on",
+    "-ctk",
+    "q4_0",
+    "-ctv",
+    "q8_0",
+    "-cram",
+    -1,
+    "-sm",
+    "layer",
+    "-dev",
+    "Vulkan1,Vulkan2",
+    "-ts",
+    "8,12",
+    "-fit",
+    "off",
+    "-mg",
+    1,
+    "--mmap",
+  ]
+
 # # Optional global model preset options applied under [*] in models.ini.
 # # Keys map to llama.cpp CLI args without leading dashes, short args, or LLAMA_ARG_* env names.
 # llama_preset_global:
diff --git a/ansible/roles/llama_server/defaults/main.yml b/ansible/roles/llama_server/defaults/main.yml
index bd295ab..b29f5e5 100644
--- a/ansible/roles/llama_server/defaults/main.yml
+++ b/ansible/roles/llama_server/defaults/main.yml
@@ -9,8 +9,10 @@ llama_server_tag: "server-vulkan"
 llama_server_extra_groups: "users,video"
 llama_server_devices: []
 llama_server_env: {}
-llama_server_port: 8080
+llama_server_port: 11434
 llama_server_preset_global: {}
+llama_server_host: "127.0.0.1"
+llama_server_argv_extra: []
 
 # Llama Exporter configuration
 llama_server_exporter_enabled: true
@@ -24,3 +26,29 @@ llama_server_exporter_scrape_timeout: 5
 
 # Llama.cpp scrape targets
 llama_server_exporter_targets: []
+
+# Windows-specific defaults
+llama_server_windows_install_dir: "C:\\llama_server"
+llama_server_windows_user: "llama"
+llama_server_windows_working_dir: "C:\\llama_server"
+llama_server_windows_version: "b9442"
+llama_server_windows_backend: "cuda"
+llama_server_cuda_version: "13.3"
+llama_server_windows_nssm_dir: "C:\\nssm"
+llama_server_windows_nssm_version: "2.24"
+llama_server_windows_nssm_url: "https://nssm.cc/release/nssm-{{ llama_server_windows_nssm_version }}.zip"
+
+# Llama-server binary defaults for Windows
+llama_server_windows_bin_name: "llama-server.exe"
+
+# Windows Python defaults
+llama_server_windows_python_version: "3.12.10"
+llama_server_windows_python_path: "C:\\Program Files\\Python{{ llama_server_windows_python_version | regex_replace('^(\\d+)\\.(\\d+)\\..*$', '\\1\\2') }}\\python.exe"
+llama_server_windows_vcredist_url: "https://aka.ms/vs/17/release/vc_redist.x64.exe"
+
+# Windows Exporter defaults
+llama_server_windows_exporter_enabled: true
+llama_server_windows_exporter_port: 9550
+llama_server_windows_exporter_interval: 15
+llama_server_windows_exporter_timeout: 5
+llama_server_windows_exporter_targets: []
diff --git a/ansible/roles/llama_server/files/scripts/llama_hf_sync_windows.py b/ansible/roles/llama_server/files/scripts/llama_hf_sync_windows.py
new file mode 100644
index 0000000..3129af3
--- /dev/null
+++ b/ansible/roles/llama_server/files/scripts/llama_hf_sync_windows.py
@@ -0,0 +1,292 @@
+#!/usr/bin/env python3
+"""Synchronize llama.cpp models from Hugging Face into a managed local directory (Windows-compatible)."""
+
+from __future__ import annotations
+
+import argparse
+import configparser
+import json
+import os
+import re
+import shutil
+import sys
+from pathlib import Path
+from typing import Any
+
+from huggingface_hub import HfApi, snapshot_download
+
+
+def _sanitize_slug(value: str) -> str:
+    slug = re.sub(r"[^a-zA-Z0-9._-]+", "-", value.strip())
+    slug = slug.strip("-._")
+    return slug.lower() or "model"
+
+
+def _resolve_repo_file(api: HfApi, model: dict[str, Any]) -> str:
+    repo_files = api.list_repo_files(repo_id=model["model_id"], revision=model["revision"], repo_type="model")
+
+    hf_file = model.get("hf_file")
+    if hf_file:
+        if hf_file in repo_files:
+            return hf_file
+        raise RuntimeError(
+            f"Requested hf_file '{hf_file}' was not found in repo "
+            f"{model['model_id']}@{model['revision']}"
+        )
+
+    quant_upper = (model.get("quant") or "").upper()
+    ggufs = [item for item in repo_files if item.lower().endswith(".gguf")]
+    matches = [item for item in ggufs if quant_upper in Path(item).name.upper()]
+    if matches:
+        return sorted(matches, key=lambda name: (len(Path(name).name), name))[0]
+
+    raise RuntimeError(
+        f"No GGUF file containing quant '{model.get('quant')}' found in "
+        f"repo {model['model_id']}@{model['revision']}. Set hf_file explicitly if needed."
+    )
+
+
+def _load_models(models_file: Path) -> list[dict[str, Any]]:
+    data = json.loads(models_file.read_text(encoding="utf-8"))
+    if not isinstance(data, list):
+        raise RuntimeError("models file must contain a JSON array")
+
+    normalized: list[dict[str, Any]] = []
+    for item in data:
+        if not isinstance(item, dict):
+            raise RuntimeError("each model entry must be an object")
+        if item.get("enable", True) is False:
+            continue
+
+        model_id = str(item.get("model_id", "")).strip()
+        quant = str(item.get("quant", "")).strip()
+        revision = str(item.get("revision", "")).strip()
+        hf_file = str(item.get("hf_file", "")).strip() or None
+        name = str(item.get("name", "")).strip()
+        preset = item.get("preset", {})
+
+        if preset is None:
+            preset = {}
+        if not isinstance(preset, dict):
+            raise RuntimeError(f"preset must be an object for model '{name or model_id or 'unknown'}'")
+
+        if not name:
+            raise RuntimeError(f"name is required for model '{model_id or 'unknown'}'")
+        if not model_id:
+            raise RuntimeError("model_id is required for all models")
+        if not quant and not hf_file:
+            raise RuntimeError(f"quant or hf_file is required for model '{model_id}'")
+        if not revision:
+            raise RuntimeError(f"revision is required for model '{model_id}'")
+
+        normalized.append(
+            {
+                "model_id": model_id,
+                "quant": quant,
+                "revision": revision,
+                "hf_file": hf_file,
+                "name": name,
+                "preset": preset,
+            }
+        )
+
+    return normalized
+
+
+def _load_preset_global(preset_global_file: Path | None) -> dict[str, Any]:
+    if preset_global_file is None:
+        return {}
+
+    if not preset_global_file.exists():
+        return {}
+
+    data = json.loads(preset_global_file.read_text(encoding="utf-8"))
+    if data is None:
+        return {}
+    if not isinstance(data, dict):
+        raise RuntimeError("global preset options must be a JSON object")
+    return data
+
+
+def _to_preset_value(value: Any) -> str:
+    if isinstance(value, bool):
+        return "true" if value else "false"
+    if isinstance(value, (int, float, str)):
+        return str(value)
+    return json.dumps(value, separators=(",", ":"))
+
+
+def _normalize_preset_options(options: dict[str, Any], scope: str) -> dict[str, str]:
+    normalized: dict[str, str] = {}
+    for key, value in options.items():
+        key_str = str(key).strip()
+        if not key_str:
+            raise RuntimeError(f"{scope} preset option keys must be non-empty")
+        if value is None:
+            continue
+        normalized[key_str] = _to_preset_value(value)
+    return normalized
+
+
+def _download_model(model: dict[str, Any], target_dir: Path, repo_file: str, dry_run: bool) -> None:
+    if dry_run:
+        return
+
+    snapshot_download(
+        repo_id=model["model_id"],
+        revision=model["revision"],
+        local_dir=str(target_dir),
+        allow_patterns=[repo_file],
+        local_dir_use_symlinks=False,
+    )
+
+
+def _write_preset(preset_file: Path, entries: list[dict[str, Any]], global_options: dict[str, Any]) -> None:
+    parser = configparser.ConfigParser(interpolation=None)
+    parser.optionxform = str
+    parser["*"] = _normalize_preset_options(global_options, "global")
+
+    for entry in entries:
+        model_options = _normalize_preset_options(entry.get("preset", {}), f"model '{entry['name']}'")
+        model_options["model"] = entry["container_model_path"]
+        parser[entry["name"]] = model_options
+
+    preset_file.parent.mkdir(parents=True, exist_ok=True)
+    with preset_file.open("w", encoding="utf-8") as fh:
+        # Router preset format supports top-level version key.
+        fh.write("version = 1\n\n")
+        parser.write(fh)
+
+
+def _prune_unmanaged(managed_dir: Path, links_dir: Path, keep_slugs: set[str], dry_run: bool) -> tuple[list[str], list[str]]:
+    pruned_dirs: list[str] = []
+    pruned_links: list[str] = []
+
+    if managed_dir.exists():
+        for child in managed_dir.iterdir():
+            if not child.is_dir():
+                continue
+            if child.name == ".router":
+                continue
+            if child.name not in keep_slugs:
+                pruned_dirs.append(child.name)
+                if not dry_run:
+                    shutil.rmtree(child)
+
+    if links_dir.exists():
+        for child in links_dir.iterdir():
+            if child.suffix != ".gguf":
+                continue
+            slug = child.stem
+            if slug not in keep_slugs:
+                pruned_links.append(child.name)
+                if not dry_run:
+                    child.unlink(missing_ok=True)
+
+    return pruned_dirs, pruned_links
+
+
+def _create_hard_link(target: Path, link_path: Path) -> None:
+    """Create a hard link on Windows, falling back to copy if hard links are not supported."""
+    if link_path.exists() or link_path.is_symlink():
+        link_path.unlink()
+
+    try:
+        # On Windows, hard links require the source to be on the same volume
+        link_path.hardlink_to(target)
+    except (OSError, NotImplementedError):
+        # Fall back to copy if hard links are not supported (different volumes, permissions, etc.)
+        shutil.copy2(str(target), str(link_path))
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser(description="Sync Hugging Face GGUF models for llama.cpp router mode (Windows)")
+    parser.add_argument("--models-file", required=True)
+    parser.add_argument("--managed-dir", required=True)
+    parser.add_argument("--links-dir", required=True)
+    parser.add_argument("--manifest-file", required=True)
+    parser.add_argument("--preset-file", required=True)
+    parser.add_argument("--preset-global-file")
+    parser.add_argument("--container-links-dir", required=True)
+    parser.add_argument("--prune", action="store_true")
+    parser.add_argument("--dry-run", action="store_true")
+    args = parser.parse_args()
+
+    models_file = Path(args.models_file)
+    managed_dir = Path(args.managed_dir)
+    links_dir = Path(args.links_dir)
+    manifest_file = Path(args.manifest_file)
+    preset_file = Path(args.preset_file)
+    preset_global_file = Path(args.preset_global_file) if args.preset_global_file else None
+
+    managed_dir.mkdir(parents=True, exist_ok=True)
+    links_dir.mkdir(parents=True, exist_ok=True)
+    manifest_file.parent.mkdir(parents=True, exist_ok=True)
+    preset_file.parent.mkdir(parents=True, exist_ok=True)
+
+    models = _load_models(models_file)
+    global_options = _load_preset_global(preset_global_file)
+    api = HfApi()
+    entries: list[dict[str, Any]] = []
+
+    for model in models:
+        slug = _sanitize_slug(model["name"])
+        target_dir = managed_dir / slug
+        target_dir.mkdir(parents=True, exist_ok=True)
+        repo_file = _resolve_repo_file(api, model)
+
+        _download_model(model, target_dir, repo_file, args.dry_run)
+        selected_file = target_dir / repo_file
+        if not args.dry_run and not selected_file.exists():
+            raise RuntimeError(f"Expected downloaded file not found: {selected_file}")
+
+        link_path = links_dir / f"{slug}.gguf"
+        container_model_path = f"{args.container_links_dir.rstrip('/')}/{slug}.gguf"
+
+        if not args.dry_run:
+            _create_hard_link(selected_file, link_path)
+
+        entries.append(
+            {
+                "name": model["name"],
+                "slug": slug,
+                "model_id": model["model_id"],
+                "revision": model["revision"],
+                "quant": model["quant"],
+                "hf_file": model.get("hf_file"),
+                "preset": model.get("preset", {}),
+                "repo_file": repo_file,
+                "selected_file": str(selected_file),
+                "link_path": str(link_path),
+                "container_model_path": container_model_path,
+            }
+        )
+
+    pruned_dirs: list[str] = []
+    pruned_links: list[str] = []
+    if args.prune:
+        keep_slugs = {entry["slug"] for entry in entries}
+        pruned_dirs, pruned_links = _prune_unmanaged(managed_dir, links_dir, keep_slugs, args.dry_run)
+
+    result = {
+        "models": entries,
+        "dry_run": args.dry_run,
+        "prune": args.prune,
+        "pruned_dirs": pruned_dirs,
+        "pruned_links": pruned_links,
+    }
+
+    if not args.dry_run:
+        manifest_file.write_text(json.dumps(result, indent=2), encoding="utf-8")
+        _write_preset(preset_file, entries, global_options)
+
+    print(json.dumps(result))
+    return 0
+
+
+if __name__ == "__main__":
+    try:
+        raise SystemExit(main())
+    except Exception as exc:
+        print(f"ERROR: {exc}", file=sys.stderr)
+        raise SystemExit(1)
diff --git a/ansible/roles/llama_server/meta/main.yml b/ansible/roles/llama_server/meta/main.yml
index 3780fd6..c7f5786 100644
--- a/ansible/roles/llama_server/meta/main.yml
+++ b/ansible/roles/llama_server/meta/main.yml
@@ -11,3 +11,7 @@ galaxy_info:
     - name: EL
       versions:
         - "9"
+    - name: Windows
+      versions:
+        - "2019"
+        - "2022"
diff --git a/ansible/roles/llama_server/tasks/llama_exporter.yml b/ansible/roles/llama_server/tasks/el/llama_exporter.yml
similarity index 100%
rename from ansible/roles/llama_server/tasks/llama_exporter.yml
rename to ansible/roles/llama_server/tasks/el/llama_exporter.yml
diff --git a/ansible/roles/llama_server/tasks/el/main.yml b/ansible/roles/llama_server/tasks/el/main.yml
new file mode 100644
index 0000000..089d719
--- /dev/null
+++ b/ansible/roles/llama_server/tasks/el/main.yml
@@ -0,0 +1,336 @@
+---
+- name: Validate llama_server_models schema
+  ansible.builtin.assert:
+    that:
+      - llama_server_models is iterable
+    fail_msg: "llama_server_models must be a list"
+
+- name: Create group
+  become: true
+  ansible.builtin.group:
+    name: "{{ llama_server_group }}"
+    system: true
+
+- name: Create user
+  become: true
+  ansible.builtin.user:
+    name: "{{ llama_server_user }}"
+    groups: "{{ podman_extra_groups | default([]) }}"
+    shell: /bin/bash
+    home: "{{ llama_server_home }}"
+    group: "{{ llama_server_group }}"
+
+- name: Disable password login for user
+  ansible.builtin.command: passwd -d "{{ llama_server_user }}"
+  become: true
+  changed_when: false
+
+- name: Create user home
+  become: true
+  ansible.builtin.file:
+    path: "{{ llama_server_home }}"
+    state: directory
+    owner: "{{ llama_server_user }}"
+    group: "{{ llama_server_group }}"
+    mode: "0750"
+
+- name: Set fact podman_binary (Debian family)
+  ansible.builtin.set_fact:
+    podman_binary: /usr/bin/podman
+  when: ansible_os_family == 'Debian'
+
+- name: Set fact podman_binary (Red Hat family)
+  ansible.builtin.set_fact:
+    podman_binary: /usr/sbin/podman
+  when: ansible_os_family == 'RedHat'
+
+- name: Create required folders
+  become: true
+  ansible.builtin.file:
+    path: "{{ llama_server_home }}/{{ item }}"
+    state: directory
+    owner: "{{ llama_server_user }}"
+    group: "{{ llama_server_group }}"
+    mode: "0750"
+  loop:
+    - models
+    - .cache
+    - .cache/llama.cpp
+
+- name: "Add SSH Authorized key"
+  become: true
+  ansible.posix.authorized_key:
+    user: "{{ llama_server_user }}"
+    key: "{{ llama_server_pubkey }}"
+    state: "{{ 'present' if llama_server_pubkey is defined and llama_server_pubkey != '' else 'absent' }}"
+  when: llama_server_pubkey is defined and llama_server_pubkey != ""
+
+- name: SELinux tasks (RedHat only)
+  when: ansible_os_family == 'RedHat'
+  become: true
+  block:
+    - name: Persistently set SELinux context
+      community.general.sefcontext:
+        target: "{{ llama_server_home }}(/.*)?"
+        setype: user_home_dir_t
+        state: present
+
+    - name: Persistently set SELinux context (ssh authorized keys)
+      community.general.sefcontext:
+        target: "{{ llama_server_home }}/.ssh/authorized_keys"
+        setype: ssh_home_t
+        state: present
+      when: llama_server_pubkey is defined and llama_server_pubkey != ""
+
+    - name: Apply SELinux context
+      ansible.builtin.command: restorecon -Rv {{ llama_server_home }}
+      changed_when: false
+
+- name: Enable sudo access to llama user (temporary)
+  become: true
+  ansible.builtin.lineinfile:
+    path: "/etc/sudoers.d/ansible_{{ llama_server_user }}_tmp"
+    create: true
+    mode: "0440"
+    line: "{{ ansible_facts['user_id'] }} ALL=({{ llama_server_user }}) NOPASSWD: ALL"
+    validate: "visudo -cf %s"
+
+- name: Enable lingering for llama user
+  become: true
+  ansible.builtin.command: loginctl enable-linger {{ llama_server_user }}
+  register: linger_status
+  changed_when: >-
+    'created' in linger_status.stdout or
+    (linger_status.rc == 0 and not
+      ('linger file already exists' in linger_status.stderr or
+      'linger file does not exist' in linger_status.stderr))
+  failed_when:
+    - linger_status.rc != 0 and not
+      ('linger file already exists' in linger_status.stderr)
+
+- name: Get current user's UID
+  ansible.builtin.command: id -u {{ llama_server_user }}
+  changed_when: false
+  register: uid
+
+- name: Get current user's GID
+  ansible.builtin.command: id -g {{ llama_server_user }}
+  changed_when: false
+  register: gid
+
+- name: Run tasks as podman user
+  become: true
+  become_user: "{{ llama_server_user }}"
+  block:
+    - name: Pull the requested images
+      containers.podman.podman_image:
+        name: "{{ llama_server_image }}"
+        tag: "{{ llama_server_tag }}"
+        state: present
+
+    - name: Create pods
+      containers.podman.podman_container:
+        name: "llama.cpp"
+        image: "{{ llama_server_image }}:{{ llama_server_tag }}"
+        state: started
+        device: "{{ llama_server_devices | default(omit) }}"
+        env: "{{ llama_server_env | default(omit) }}"
+        ports: "0.0.0.0:{{ llama_server_port }}:8080/tcp"
+        volumes:
+          - "{{ llama_server_home }}/models:/models:Z"
+          - "{{ llama_server_home }}/.cache:/app/.cache:Z"
+        command: >-
+          {{
+            [
+              '--metrics',
+              '--models-preset',
+              '/models/.router/models.ini',
+              '--models-max',
+              (llama_server_models_max | string)
+            ]
+            + (llama_server_argv_extra | default([]))
+          }}
+
+        cmd_args:
+          - "--userns=keep-id"
+          - "--security-opt=label=disable"
+
+- name: Create systemd service file for pods
+  become: true
+  ansible.builtin.template:
+    src: podman.service.j2
+    dest: "/etc/systemd/system/podman-llama.cpp.service"
+    owner: root
+    group: root
+    mode: "0644"
+
+- name: Reload systemd daemon
+  become: true
+  ansible.builtin.systemd:
+    daemon_reload: true
+
+- name: Enable containers at boot
+  become: true
+  ansible.builtin.systemd:
+    name: "podman-llama.cpp"
+    enabled: true
+    state: "started"
+
+- name: Install Python packages required for llama server model sync
+  become: true
+  ansible.builtin.package:
+    name: "{{ __llama_sync_system_packages__ }}"
+    state: present
+  vars:
+    __llama_sync_system_packages__: >-
+      {{
+        {
+          'Debian': ['python3-pip', 'python3-packaging', 'python3-venv'],
+          'RedHat': ['python3-pip', 'python3-packaging'],
+          'Archlinux': ['python-pip', 'python-packaging']
+        }[ansible_os_family]
+      }}
+
+- name: Install Hugging Face hub client in llama server sync virtualenv
+  become: true
+  become_user: "{{ llama_server_user }}"
+  ansible.builtin.pip:
+    name:
+      - pip
+      - setuptools
+      - packaging
+      - huggingface_hub>=0.32.0
+    state: present
+    virtualenv: "{{ llama_server_home }}/.router/.venv"
+    virtualenv_command: "python3 -m venv"
+
+- name: Create llama server router model directories
+  become: true
+  ansible.builtin.file:
+    path: "{{ item.path }}"
+    state: directory
+    owner: "{{ llama_server_user }}"
+    group: "{{ llama_server_group }}"
+    mode: "{{ item.mode }}"
+  loop:
+    - { path: "{{ llama_server_home }}/models", mode: "0750" }
+    - { path: "{{ llama_server_home }}/models/managed", mode: "0750" }
+    - { path: "{{ llama_server_home }}/models/managed-links", mode: "0750" }
+    - { path: "{{ llama_server_home }}/models/.router", mode: "0750" }
+
+- name: Copy llama server Hugging Face sync helper
+  become: true
+  ansible.builtin.copy:
+    src: scripts/llama_hf_sync.py
+    dest: "{{ llama_server_home }}/models/.router/llama_hf_sync.py"
+    owner: "{{ llama_server_user }}"
+    group: "{{ llama_server_group }}"
+    mode: "0750"
+
+- name: Write desired llama server model manifest input
+  become: true
+  ansible.builtin.copy:
+    content: "{{ llama_server_models | to_nice_json }}"
+    dest: "{{ llama_server_home }}/models/.router/desired-models.json"
+    owner: "{{ llama_server_user }}"
+    group: "{{ llama_server_group }}"
+    mode: "0640"
+
+- name: Write global llama server preset options
+  become: true
+  ansible.builtin.copy:
+    content: "{{ llama_server_preset_global | to_nice_json }}"
+    dest: "{{ llama_server_home }}/models/.router/preset-global.json"
+    owner: "{{ llama_server_user }}"
+    group: "{{ llama_server_group }}"
+    mode: "0640"
+
+- name: Set arguments for llama server model sync
+  ansible.builtin.set_fact:
+    __llama_preset_file__:
+      - "{{ llama_server_home }}/.router/.venv/bin/python"
+      - "{{ llama_server_home }}/models/.router/llama_hf_sync.py"
+      - "--models-file"
+      - "{{ llama_server_home }}/models/.router/desired-models.json"
+      - "--managed-dir"
+      - "{{ llama_server_home }}/models/managed"
+      - "--links-dir"
+      - "{{ llama_server_home }}/models/managed-links"
+      - "--manifest-file"
+      - "{{ llama_server_home }}/models/.router/manifest.json"
+      - "--preset-file"
+      - "{{ llama_server_home }}/models/.router/models.ini"
+      - "--preset-global-file"
+      - "{{ llama_server_home }}/models/.router/preset-global.json"
+      - "--container-links-dir"
+      - "/models/managed-links"
+
+- name: Run llama server model sync from Hugging Face
+  vars:
+    __llama_sync_argv__: >-
+      {{ __llama_preset_file__
+        + (['--prune'] if llama_server_models_prune_enabled | default(true) else [])
+        + (['--dry-run'] if llama_server_models_dry_run | default(false) else [])
+      }}
+  become: true
+  become_user: "{{ llama_server_user }}"
+  ansible.builtin.command:
+    argv: "{{ __llama_sync_argv__ }}"
+  environment:
+    HF_TOKEN: "{{ lookup('env', 'HF_TOKEN') | default('', true) }}"
+  register: __llama_sync_output__
+  changed_when: false
+
+- name: Parse llama server sync output
+  ansible.builtin.set_fact:
+    __llama_sync_result__: "{{ __llama_sync_output__.stdout | from_json }}"
+
+- name: Stat llama router preset file (when not in dry-run mode)
+  become: true
+  become_user: "{{ llama_server_user }}"
+  ansible.builtin.stat:
+    path: "{{ llama_server_home }}/models/.router/models.ini"
+  register: __llama_preset_stat__
+  when: not (llama_server_models_dry_run | default(false))
+
+- name: Ensure router preset file exists when not in dry-run mode
+  ansible.builtin.assert:
+    that:
+      - __llama_preset_stat__.stat.exists
+    fail_msg: "llama router preset file was not created"
+  when: not (llama_server_models_dry_run | default(false))
+
+- name: Query llama router model list
+  ansible.builtin.uri:
+    url: "http://127.0.0.1:{{ llama_server_port | default(11434) }}/models?reload=1"
+    method: GET
+    status_code: 200
+  register: __llama_router_models__
+  changed_when: false
+  when:
+    - llama_server_models_sync_enabled | default(true)
+    - not (llama_server_models_dry_run | default(false))
+
+- name: Validate expected model aliases are available in router
+  ansible.builtin.assert:
+    that:
+      - item.name in (__llama_router_models__.json.data | map(attribute='id') | list)
+    fail_msg: "Model name '{{ item.name }}' not present in llama router /models output"
+  loop: "{{ __llama_sync_result__.models | default([]) }}"
+  loop_control:
+    label: "{{ item.name }}"
+  when:
+    - llama_server_models_sync_enabled | default(true)
+    - not (llama_server_models_dry_run | default(false))
+
+- name: Include Llama Exporter setup
+  when: llama_server_exporter_enabled | default(true) | bool
+  ansible.builtin.include_tasks: llama_exporter.yml
+  tags:
+    - roles::llama_server::llama_exporter
+
+- name: Remove temporary sudo access
+  become: true
+  ansible.builtin.file:
+    path: "/etc/sudoers.d/ansible_{{ llama_server_user }}_tmp"
+    state: absent
diff --git a/ansible/roles/llama_server/tasks/main.yml b/ansible/roles/llama_server/tasks/main.yml
index 3e0e3dd..8475123 100644
--- a/ansible/roles/llama_server/tasks/main.yml
+++ b/ansible/roles/llama_server/tasks/main.yml
@@ -1,358 +1,17 @@
 ---
-- name: Validate llama_server_models schema
-  ansible.builtin.assert:
-    that:
-      - llama_server_models is iterable
-    fail_msg: "llama_server_models must be a list"
+# --- EL-specific setup (podman + systemd) ---
 
-- name: Create group
-  become: true
-  ansible.builtin.group:
-    name: "{{ llama_server_group }}"
-    system: true
-
-- name: Create user
-  become: true
-  ansible.builtin.user:
-    name: "{{ llama_server_user }}"
-    groups: "{{ podman_extra_groups | default([]) }}"
-    shell: /bin/bash
-    home: "{{ llama_server_home }}"
-    group: "{{ llama_server_group }}"
-
-- name: Disable password login for user
-  ansible.builtin.command: passwd -d "{{ llama_server_user }}"
-  become: true
-  changed_when: false
-
-- name: Create user home
-  become: true
-  ansible.builtin.file:
-    path: "{{ llama_server_home }}"
-    state: directory
-    owner: "{{ llama_server_user }}"
-    group: "{{ llama_server_group }}"
-    mode: "0750"
-
-- name: Set fact podman_binary (Debian family)
-  ansible.builtin.set_fact:
-    podman_binary: /usr/bin/podman
-  when: ansible_os_family == 'Debian'
-
-- name: Set fact podman_binary (Red Hat family)
-  ansible.builtin.set_fact:
-    podman_binary: /usr/sbin/podman
+- name: Run EL container setup
   when: ansible_os_family == 'RedHat'
-
-- name: Create required folders
-  become: true
-  ansible.builtin.file:
-    path: "{{ llama_server_home }}/{{ item }}"
-    state: directory
-    owner: "{{ llama_server_user }}"
-    group: "{{ llama_server_group }}"
-    mode: "0750"
-  loop:
-    - models
-    - .cache
-    - .cache/llama.cpp
-
-- name: "Add SSH Authorized key"
-  become: true
-  ansible.posix.authorized_key:
-    user: "{{ llama_server_user }}"
-    key: "{{ llama_server_pubkey }}"
-    state: "{{ 'present' if llama_server_pubkey is defined and llama_server_pubkey != '' else 'absent' }}"
-  when: llama_server_pubkey is defined and llama_server_pubkey != ""
-
-- name: SELinux tasks (RedHat only)
-  when: ansible_os_family == 'RedHat'
-  become: true
-  block:
-    - name: Persistently set SELinux context
-      community.general.sefcontext:
-        target: "{{ llama_server_home }}(/.*)?"
-        setype: user_home_dir_t
-        state: present
-
-    - name: Persistently set SELinux context (ssh authorized keys)
-      community.general.sefcontext:
-        target: "{{ llama_server_home }}/.ssh/authorized_keys"
-        setype: ssh_home_t
-        state: present
-      when: llama_server_pubkey is defined and llama_server_pubkey != ""
-
-    - name: Apply SELinux context
-      ansible.builtin.command: restorecon -Rv {{ llama_server_home }}
-      changed_when: false
-
-- name: Enable sudo access to llama user (temporary)
-  become: true
-  ansible.builtin.lineinfile:
-    path: "/etc/sudoers.d/ansible_{{ llama_server_user }}_tmp"
-    create: true
-    mode: "0440"
-    line: "{{ ansible_facts['user_id'] }} ALL=({{ llama_server_user }}) NOPASSWD: ALL"
-    validate: "visudo -cf %s"
-
-- name: Enable lingering for llama user
-  become: true
-  ansible.builtin.command: loginctl enable-linger {{ llama_server_user }}
-  register: linger_status
-  changed_when: >-
-    'created' in linger_status.stdout or
-    (linger_status.rc == 0 and not
-      ('linger file already exists' in linger_status.stderr or
-      'linger file does not exist' in linger_status.stderr))
-  failed_when:
-    - linger_status.rc != 0 and not
-      ('linger file already exists' in linger_status.stderr)
-
-- name: Get current user's UID
-  ansible.builtin.command: id -u {{ llama_server_user }}
-  changed_when: false
-  register: uid
-
-- name: Get current user's GID
-  ansible.builtin.command: id -g {{ llama_server_user }}
-  changed_when: false
-  register: gid
-
-- name: Run tasks as podman user
-  become: true
-  become_user: "{{ llama_server_user }}"
-  block:
-    - name: Pull the requested images
-      containers.podman.podman_image:
-        name: "{{ llama_server_image }}"
-        tag: "{{ llama_server_tag }}"
-        state: present
-
-    - name: Create pods
-      containers.podman.podman_container:
-        name: "llama.cpp"
-        image: "{{ llama_server_image }}:{{ llama_server_tag }}"
-        state: started
-        device: "{{ llama_server_devices | default(omit) }}"
-        env: "{{ llama_server_env | default(omit) }}"
-        ports: "0.0.0.0:{{ llama_server_port }}:8080/tcp"
-        volumes:
-          - "{{ llama_server_home }}/models:/models:Z"
-          - "{{ llama_server_home }}/.cache:/app/.cache:Z"
-        command:
-          - "-t"
-          - "{{ llama_server_argv_threads | default(10) }}"
-          - "-np"
-          - "{{ llama_server_argv_parallel | default(1) }}"
-          - "-b"
-          - "{{ llama_server_argv_batch_size | default(1024) }}"
-          - "-ub"
-          - "{{ llama_server_argv_ubatch_size | default(512) }}"
-          - "-fa"
-          - "{{ llama_server_argv_flash_attn | default('on') }}"
-          - "-ctk"
-          - "{{ llama_server_argv_cache_type_k | default('q4_0') }}"
-          - "-ctv"
-          - "{{ llama_server_argv_cache_type_v | default('q8_0') }}"
-          - "-cram"
-          - "{{ llama_server_argv_cache_reuse | default(-1) }}"
-          - "-sm"
-          - "{{ llama_server_argv_split_mode | default('layer') }}"
-          - "-dev"
-          - "{{ llama_server_argv_devices | default('Vulkan1,Vulkan2') }}"
-          - "-ts"
-          - "{{ llama_server_argv_tensor_split | default('8,12') }}"
-          - "-fit"
-          - "{{ llama_server_argv_fit | default('off') }}"
-          - "-mg"
-          - "{{ llama_server_argv_main_gpu | default(1) }}"
-          - "--mmap"
-          - "--metrics"
-          - "--models-preset"
-          - "/models/.router/models.ini"
-          - "--models-max"
-          - "{{ llama_server_models_max }}"
-
-        cmd_args:
-          - "--userns=keep-id"
-          - "--security-opt=label=disable"
-
-- name: Create systemd service file for pods
-  become: true
-  ansible.builtin.template:
-    src: podman.service.j2
-    dest: "/etc/systemd/system/podman-llama.cpp.service"
-    owner: root
-    group: root
-    mode: "0644"
-
-- name: Reload systemd daemon
-  become: true
-  ansible.builtin.systemd:
-    daemon_reload: true
-
-- name: Enable containers at boot
-  become: true
-  ansible.builtin.systemd:
-    name: "podman-llama.cpp"
-    enabled: true
-    state: "started"
-
-- name: Install Python packages required for llama server model sync
-  become: true
-  ansible.builtin.package:
-    name: "{{ __llama_sync_system_packages__ }}"
-    state: present
-  vars:
-    __llama_sync_system_packages__: >-
-      {{
-        {
-          'Debian': ['python3-pip', 'python3-packaging', 'python3-venv'],
-          'RedHat': ['python3-pip', 'python3-packaging'],
-          'Archlinux': ['python-pip', 'python-packaging']
-        }[ansible_os_family]
-      }}
-
-- name: Install Hugging Face hub client in llama server sync virtualenv
-  become: true
-  become_user: "{{ llama_server_user }}"
-  ansible.builtin.pip:
-    name:
-      - pip
-      - setuptools
-      - packaging
-      - huggingface_hub>=0.32.0
-    state: present
-    virtualenv: "{{ llama_server_home }}/.router/.venv"
-    virtualenv_command: "python3 -m venv"
-
-- name: Create llama server router model directories
-  become: true
-  ansible.builtin.file:
-    path: "{{ item.path }}"
-    state: directory
-    owner: "{{ llama_server_user }}"
-    group: "{{ llama_server_group }}"
-    mode: "{{ item.mode }}"
-  loop:
-    - { path: "{{ llama_server_home }}/models", mode: "0750" }
-    - { path: "{{ llama_server_home }}/models/managed", mode: "0750" }
-    - { path: "{{ llama_server_home }}/models/managed-links", mode: "0750" }
-    - { path: "{{ llama_server_home }}/models/.router", mode: "0750" }
-
-- name: Copy llama server Hugging Face sync helper
-  become: true
-  ansible.builtin.copy:
-    src: scripts/llama_hf_sync.py
-    dest: "{{ llama_server_home }}/models/.router/llama_hf_sync.py"
-    owner: "{{ llama_server_user }}"
-    group: "{{ llama_server_group }}"
-    mode: "0750"
-
-- name: Write desired llama server model manifest input
-  become: true
-  ansible.builtin.copy:
-    content: "{{ llama_server_models | to_nice_json }}"
-    dest: "{{ llama_server_home }}/models/.router/desired-models.json"
-    owner: "{{ llama_server_user }}"
-    group: "{{ llama_server_group }}"
-    mode: "0640"
-
-- name: Write global llama server preset options
-  become: true
-  ansible.builtin.copy:
-    content: "{{ llama_server_preset_global | to_nice_json }}"
-    dest: "{{ llama_server_home }}/models/.router/preset-global.json"
-    owner: "{{ llama_server_user }}"
-    group: "{{ llama_server_group }}"
-    mode: "0640"
-
-- name: Set arguments for llama server model sync
-  ansible.builtin.set_fact:
-    __llama_preset_file__:
-      - "{{ llama_server_home }}/.router/.venv/bin/python"
-      - "{{ llama_server_home }}/models/.router/llama_hf_sync.py"
-      - "--models-file"
-      - "{{ llama_server_home }}/models/.router/desired-models.json"
-      - "--managed-dir"
-      - "{{ llama_server_home }}/models/managed"
-      - "--links-dir"
-      - "{{ llama_server_home }}/models/managed-links"
-      - "--manifest-file"
-      - "{{ llama_server_home }}/models/.router/manifest.json"
-      - "--preset-file"
-      - "{{ llama_server_home }}/models/.router/models.ini"
-      - "--preset-global-file"
-      - "{{ llama_server_home }}/models/.router/preset-global.json"
-      - "--container-links-dir"
-      - "/models/managed-links"
-
-- name: Run llama server model sync from Hugging Face
-  vars:
-    __llama_sync_argv__: >-
-      {{ __llama_preset_file__
-        + (['--prune'] if llama_server_models_prune_enabled | default(true) else [])
-        + (['--dry-run'] if llama_server_models_dry_run | default(false) else [])
-      }}
-  become: true
-  become_user: "{{ llama_server_user }}"
-  ansible.builtin.command:
-    argv: "{{ __llama_sync_argv__ }}"
-  environment:
-    HF_TOKEN: "{{ lookup('env', 'HF_TOKEN') | default('', true) }}"
-  register: __llama_sync_output__
-  changed_when: false
-
-- name: Parse llama server sync output
-  ansible.builtin.set_fact:
-    __llama_sync_result__: "{{ __llama_sync_output__.stdout | from_json }}"
-
-- name: Stat llama router preset file (when not in dry-run mode)
-  become: true
-  become_user: "{{ llama_server_user }}"
-  ansible.builtin.stat:
-    path: "{{ llama_server_home }}/models/.router/models.ini"
-  register: __llama_preset_stat__
-  when: not (llama_server_models_dry_run | default(false))
-
-- name: Ensure router preset file exists when not in dry-run mode
-  ansible.builtin.assert:
-    that:
-      - __llama_preset_stat__.stat.exists
-    fail_msg: "llama router preset file was not created"
-  when: not (llama_server_models_dry_run | default(false))
-
-- name: Query llama router model list
-  ansible.builtin.uri:
-    url: "http://127.0.0.1:{{ llama_server_port | default(8080) }}/models?reload=1"
-    method: GET
-    status_code: 200
-  register: __llama_router_models__
-  changed_when: false
-  when:
-    - llama_server_models_sync_enabled | default(true)
-    - not (llama_server_models_dry_run | default(false))
-
-- name: Validate expected model aliases are available in router
-  ansible.builtin.assert:
-    that:
-      - item.name in (__llama_router_models__.json.data | map(attribute='id') | list)
-    fail_msg: "Model name '{{ item.name }}' not present in llama router /models output"
-  loop: "{{ __llama_sync_result__.models | default([]) }}"
-  loop_control:
-    label: "{{ item.name }}"
-  when:
-    - llama_server_models_sync_enabled | default(true)
-    - not (llama_server_models_dry_run | default(false))
-
-- name: Include Llama Exporter setup
-  when: llama_server_exporter_enabled | default(true) | bool
-  ansible.builtin.include_tasks: llama_exporter.yml
+  ansible.builtin.include_tasks:
+    file: el/main.yml
   tags:
-    - roles::llama_server::llama_exporter
+    - roles::llama_server::el
 
-- name: Remove temporary sudo access
-  become: true
-  ansible.builtin.file:
-    path: "/etc/sudoers.d/ansible_{{ llama_server_user }}_tmp"
-    state: absent
+# --- Windows-specific setup (NSSM) ---
+- name: Run Windows setup
+  when: ansible_os_family == 'Windows'
+  ansible.builtin.include_tasks:
+    file: windows/main.yml
+  tags:
+    - roles::llama_server::windows
diff --git a/ansible/roles/llama_server/tasks/windows/llama_exporter.yml b/ansible/roles/llama_server/tasks/windows/llama_exporter.yml
new file mode 100644
index 0000000..641f5fb
--- /dev/null
+++ b/ansible/roles/llama_server/tasks/windows/llama_exporter.yml
@@ -0,0 +1,163 @@
+---
+# Windows Exporter tasks for llama_server role
+
+- name: Define Windows Exporter paths
+  ansible.builtin.set_fact:
+    __exporter_install_dir: "{{ llama_server_windows_install_dir }}\\exporter"
+    __exporter_script: "{{ llama_server_windows_install_dir }}\\exporter\\llama_exporter.py"
+    __exporter_venv_path: "{{ llama_server_windows_install_dir }}\\exporter\\.venv"
+    __exporter_venv_python: "{{ llama_server_windows_install_dir }}\\exporter\\.venv\\Scripts\\python.exe"
+    __exporter_env: "LLAMA_EXPORTER_PORT={{ llama_server_windows_exporter_port | default(9550) }}"
+    __exporter_interval: "{{ llama_server_windows_exporter_interval | default(15) }}"
+    __exporter_timeout: "{{ llama_server_windows_exporter_timeout | default(5) }}"
+
+- name: Create exporter directories
+  ansible.windows.win_file:
+    path: "{{ item }}"
+    state: directory
+  loop:
+    - "{{ llama_server_windows_install_dir }}\\exporter"
+
+- name: Copy Windows Exporter script
+  ansible.windows.win_copy:
+    src: scripts/llama_exporter.py
+    dest: "{{ __exporter_script }}"
+
+- name: Install Python packages for Windows Exporter
+  ansible.windows.win_stat:
+    path: "{{ __exporter_venv_python }}"
+  register: __exporter_venv_python_check
+
+- name: Create exporter virtualenv
+  ansible.windows.win_command:
+    argv:
+      - "{{ llama_server_windows_python_path }}"
+      - -m
+      - venv
+      - "{{ __exporter_venv_path }}"
+  changed_when: false
+  when: not __exporter_venv_python_check.stat.exists
+
+- name: Install Python packages for Windows Exporter
+  ansible.windows.win_command:
+    argv:
+      - "{{ __exporter_venv_python }}"
+      - -m
+      - pip
+      - install
+      - --upgrade
+      - pip
+      - setuptools
+  changed_when: false
+
+- name: Remove existing exporter service if present
+  ansible.windows.win_service:
+    name: llama-exporter
+    state: absent
+  register: __exporter_remove_service
+  failed_when: false
+
+- name: Set exporter targets environment variable if provided
+  ansible.builtin.set_fact:
+    __exporter_targets_env: "LLAMA_TARGETS={{ llama_server_windows_exporter_targets | to_json }}"
+  when: llama_server_windows_exporter_targets | default([]) | length > 0
+
+- name: Install Windows Exporter service via NSSM
+  ansible.windows.win_command:
+    argv:
+      - "{{ __nssm_exe }}"
+      - install
+      - llama-exporter
+      - "{{ __exporter_venv_python }}"
+      - "{{ __exporter_script }}"
+  args:
+    chdir: "{{ __exporter_install_dir }}"
+  register: __exporter_nssm_install
+  changed_when: __exporter_nssm_install.rc == 0
+
+- name: Set exporter AppDirectory via NSSM
+  ansible.windows.win_command:
+    argv:
+      - "{{ __nssm_exe }}"
+      - set
+      - llama-exporter
+      - AppDirectory
+      - "{{ __exporter_install_dir }}"
+
+- name: Configure exporter stdout log via NSSM
+  ansible.windows.win_command:
+    argv:
+      - "{{ __nssm_exe }}"
+      - set
+      - llama-exporter
+      - AppStdout
+      - "{{ llama_server_windows_install_dir }}\\exporter-service.out.log"
+
+- name: Configure exporter stderr log via NSSM
+  ansible.windows.win_command:
+    argv:
+      - "{{ __nssm_exe }}"
+      - set
+      - llama-exporter
+      - AppStderr
+      - "{{ llama_server_windows_install_dir }}\\exporter-service.err.log"
+
+- name: Set exporter environment variables
+  ansible.windows.win_command:
+    argv:
+      - "{{ __nssm_exe }}"
+      - set
+      - llama-exporter
+      - AppEnvironmentExtra
+      - "{{ (__exporter_env ~ (('\n' ~ __exporter_targets_env) if (__exporter_targets_env is defined) else '')) }}"
+  when:
+    - __exporter_env | default('') != ''
+
+- name: Set exporter Start mode to AUTO via NSSM
+  ansible.windows.win_command:
+    argv:
+      - "{{ __nssm_exe }}"
+      - set
+      - llama-exporter
+      - Start
+      - SERVICE_AUTO_START
+
+- name: Ensure Windows Exporter service is running
+  ansible.windows.win_service:
+    name: llama-exporter
+    start_mode: auto
+    state: restarted
+  when: not ansible_check_mode | default(false)
+
+- name: Wait for exporter to listen on port
+  ansible.windows.win_wait_for:
+    port: "{{ llama_server_windows_exporter_port | default(9550) }}"
+    delay: 5
+    timeout: 30
+  when: not ansible_check_mode | default(false)
+
+- name: Set firewall remote IPs when string provided
+  ansible.builtin.set_fact:
+    llama_server_windows_remoteip: "{{ llama_server_windows_sources }}"
+  when:
+    - llama_server_windows_sources is defined
+    - llama_server_windows_sources is string
+
+- name: Set firewall remote IPs when list provided
+  ansible.builtin.set_fact:
+    llama_server_windows_remoteip: "{{ llama_server_windows_sources | join(',') }}"
+  when:
+    - llama_server_windows_sources is defined
+    - not (llama_server_windows_sources is string)
+
+- name: Allow exporter port through Windows Firewall
+  community.windows.win_firewall_rule:
+    name: Llama Exporter "{{ llama_server_windows_exporter_port | default(9550) }}"
+    localport: "{{ llama_server_windows_exporter_port | default(9550) }}"
+    protocol: tcp
+    direction: in
+    action: allow
+    enabled: true
+    # Provide one or more IPs/CIDRs in llama_server_windows_sources (e.g. ['203.0.113.10'])
+    remoteip: "{{ llama_server_windows_remoteip | default(omit) }}"
+    state: present
diff --git a/ansible/roles/llama_server/tasks/windows/main.yml b/ansible/roles/llama_server/tasks/windows/main.yml
new file mode 100644
index 0000000..6e568e8
--- /dev/null
+++ b/ansible/roles/llama_server/tasks/windows/main.yml
@@ -0,0 +1,382 @@
+---
+# Windows-specific setup for llama_server role
+
+- name: Define Windows paths and versions
+  ansible.builtin.set_fact:
+    __llama_user: "{{ llama_server_windows_user }}"
+    __nssm_exe: "{{ llama_server_windows_nssm_dir }}\\nssm.exe"
+    __llama_release_asset_name: "{{ 'llama-' ~ llama_server_windows_version ~ '-bin-win-' ~ llama_server_windows_backend ~ (llama_server_windows_backend == 'cuda') | ternary('-' ~ llama_server_cuda_version, '') ~ '-x64.zip' }}"
+    __llama_release_zip_url: "https://github.com/ggml-org/llama.cpp/releases/download/{{ llama_server_windows_version }}/{{ 'llama-' ~ llama_server_windows_version ~ '-bin-win-' ~ llama_server_windows_backend ~ (llama_server_windows_backend == 'cuda') | ternary('-' ~ llama_server_cuda_version, '') ~ '-x64.zip' }}"
+    __llama_release_zip_path: "{{ (ansible_env.TEMP | default('C:\\Windows\\Temp')) }}\\{{ 'llama-' ~ llama_server_windows_version ~ '-bin-win-' ~ llama_server_windows_backend ~ (llama_server_windows_backend == 'cuda') | ternary('-' ~ llama_server_cuda_version, '') ~ '-x64.zip' }}"
+    __llama_release_marker: "{{ llama_server_windows_install_dir }}\\.release-{{ 'llama-' ~ llama_server_windows_version ~ '-bin-win-' ~ llama_server_windows_backend ~ (llama_server_windows_backend == 'cuda') | ternary('-' ~ llama_server_cuda_version, '') ~ '-x64.zip' }}"
+    __nssm_zip_path: "{{ (ansible_env.TEMP | default('C:\\Windows\\Temp')) }}\\nssm.zip"
+    __llama_sync_script: "{{ llama_server_windows_working_dir }}\\.router\\llama_hf_sync_windows.py"
+    __llama_models_json: "{{ llama_server_windows_working_dir }}\\.router\\desired-models.json"
+    __llama_preset_global_json: "{{ llama_server_windows_working_dir }}\\.router\\preset-global.json"
+    __llama_managed_dir: "{{ llama_server_windows_working_dir }}\\models\\managed"
+    __llama_links_dir: "{{ llama_server_windows_working_dir }}\\models\\managed-links"
+    __llama_router_dir: "{{ llama_server_windows_working_dir }}\\.router"
+    __llama_manifest: "{{ llama_server_windows_working_dir }}\\.router\\manifest.json"
+    __llama_preset_file: "{{ llama_server_windows_working_dir }}\\.router\\models.ini"
+    __llama_venv_path: "{{ llama_server_windows_working_dir }}\\.router\\.venv"
+    __llama_venv_python: "{{ llama_server_windows_working_dir }}\\.router\\.venv\\Scripts\\python.exe"
+
+- name: Validate llama_server_models schema
+  ansible.builtin.assert:
+    that:
+      - llama_server_models is iterable
+    fail_msg: "llama_server_models must be a list"
+
+- name: Create llama server installation directories
+  ansible.windows.win_file:
+    path: "{{ item }}"
+    state: directory
+  loop:
+    - "{{ llama_server_windows_install_dir }}"
+    - "{{ llama_server_windows_working_dir }}"
+    - "{{ llama_server_windows_working_dir }}\\models\\managed"
+    - "{{ llama_server_windows_working_dir }}\\models\\managed-links"
+    - "{{ llama_server_windows_working_dir }}\\.router"
+
+- name: Check if Python is available
+  ansible.windows.win_stat:
+    path: "{{ llama_server_windows_python_path }}"
+  register: __llama_python_check
+
+- name: Download Python installer
+  ansible.windows.win_get_url:
+    url: "https://www.python.org/ftp/python/{{ llama_server_windows_python_version }}/python-{{ llama_server_windows_python_version }}-amd64.exe"
+    dest: "{{ (ansible_env.TEMP | default('C:\\Windows\\Temp')) }}\\python-{{ llama_server_windows_python_version }}-amd64.exe"
+    force: false
+  when: not __llama_python_check.stat.exists
+
+- name: Install Python if not available
+  ansible.windows.win_package:
+    path: "{{ (ansible_env.TEMP | default('C:\\Windows\\Temp')) }}\\python-{{ llama_server_windows_python_version }}-amd64.exe"
+    arguments: /quiet InstallAllUsers=1 PrependPath=0 Include_test=0
+    state: present
+  when: not __llama_python_check.stat.exists
+
+- name: Check Microsoft VC++ runtime availability
+  ansible.windows.win_stat:
+    path: "C:\\Windows\\System32\\vcruntime140.dll"
+  register: __llama_vcredist_check
+
+- name: Download Microsoft VC++ redistributable
+  ansible.windows.win_get_url:
+    url: "{{ llama_server_windows_vcredist_url }}"
+    dest: "{{ (ansible_env.TEMP | default('C:\\Windows\\Temp')) }}\\vc_redist.x64.exe"
+    force: false
+  when: not __llama_vcredist_check.stat.exists
+
+- name: Install Microsoft VC++ redistributable
+  ansible.windows.win_package:
+    path: "{{ (ansible_env.TEMP | default('C:\\Windows\\Temp')) }}\\vc_redist.x64.exe"
+    arguments: /install /quiet /norestart
+    state: present
+  when: not __llama_vcredist_check.stat.exists
+
+- name: Download llama.cpp Windows binary
+  ansible.windows.win_get_url:
+    url: "{{ __llama_release_zip_url }}"
+    dest: "{{ __llama_release_zip_path }}"
+    force: false
+
+- name: Check deployed llama.cpp release marker
+  ansible.windows.win_stat:
+    path: "{{ __llama_release_marker }}"
+  register: __llama_release_marker_stat
+
+- name: Remove stale llama-server binary before release upgrade
+  ansible.windows.win_file:
+    path: "{{ llama_server_windows_install_dir }}\\{{ llama_server_windows_bin_name }}"
+    state: absent
+  when: not __llama_release_marker_stat.stat.exists
+
+- name: Extract llama.cpp binary
+  community.windows.win_unzip:
+    src: "{{ __llama_release_zip_path }}"
+    dest: "{{ llama_server_windows_install_dir }}"
+    delete_archive: false
+  when: not __llama_release_marker_stat.stat.exists
+
+- name: Mark deployed llama.cpp release
+  ansible.windows.win_copy:
+    content: "{{ __llama_release_asset_name }}"
+    dest: "{{ __llama_release_marker }}"
+  when: not __llama_release_marker_stat.stat.exists
+
+- name: Verify llama-server binary is present
+  ansible.windows.win_stat:
+    path: "{{ llama_server_windows_install_dir }}\\{{ llama_server_windows_bin_name }}"
+  register: __llama_bin_stat
+
+- name: Fail if llama-server binary is missing
+  ansible.builtin.fail:
+    msg: >-
+      llama-server binary not found at {{ llama_server_windows_install_dir }}\{{ llama_server_windows_bin_name }}.
+      Check download/unzip and llama_server_windows_backend settings.
+  when: not __llama_bin_stat.stat.exists
+
+- name: Check if NSSM is installed
+  ansible.windows.win_stat:
+    path: "{{ llama_server_windows_nssm_dir }}\\nssm-{{ llama_server_windows_nssm_version }}\\win64\\nssm.exe"
+  register: __nssm_stat
+
+- name: Download NSSM if not installed
+  ansible.windows.win_get_url:
+    url: "{{ llama_server_windows_nssm_url }}"
+    dest: "{{ __nssm_zip_path }}"
+    force: false
+  when: not __nssm_stat.stat.exists
+
+- name: Extract NSSM if not installed
+  community.windows.win_unzip:
+    src: "{{ __nssm_zip_path }}"
+    dest: "{{ llama_server_windows_nssm_dir }}"
+    delete_archive: false
+    creates: "{{ llama_server_windows_nssm_dir }}\\nssm-{{ llama_server_windows_nssm_version }}\\win64\\nssm.exe"
+  when: not __nssm_stat.stat.exists
+
+- name: Copy NSSM executable
+  ansible.windows.win_copy:
+    src: "{{ llama_server_windows_nssm_dir }}\\nssm-{{ llama_server_windows_nssm_version }}\\win64\\nssm.exe"
+    dest: "{{ __nssm_exe }}"
+    remote_src: true
+
+- name: Copy Windows-compatible Hugging Face sync script
+  ansible.windows.win_copy:
+    src: scripts/llama_hf_sync_windows.py
+    dest: "{{ __llama_sync_script }}"
+    mode: "0750"
+
+- name: Write desired llama server model manifest input
+  ansible.windows.win_copy:
+    content: "{{ llama_server_models | to_nice_json }}"
+    dest: "{{ __llama_models_json }}"
+
+- name: Write global llama server preset options
+  ansible.windows.win_copy:
+    content: "{{ llama_server_preset_global | to_nice_json }}"
+    dest: "{{ __llama_preset_global_json }}"
+
+- name: Check llama sync virtualenv
+  ansible.windows.win_stat:
+    path: "{{ __llama_venv_python }}"
+  register: __llama_venv_python_check
+
+- name: Create llama sync virtualenv
+  ansible.windows.win_command:
+    argv:
+      - "{{ llama_server_windows_python_path }}"
+      - -m
+      - venv
+      - "{{ __llama_venv_path }}"
+  changed_when: false
+  when: not __llama_venv_python_check.stat.exists
+
+- name: Upgrade pip tooling in llama sync virtualenv
+  ansible.windows.win_command:
+    argv:
+      - "{{ __llama_venv_python }}"
+      - -m
+      - pip
+      - install
+      - --upgrade
+      - pip
+      - setuptools
+      - packaging
+  changed_when: false
+
+- name: Install Hugging Face dependency in llama sync virtualenv
+  ansible.windows.win_command:
+    argv:
+      - "{{ __llama_venv_python }}"
+      - -m
+      - pip
+      - install
+      - "huggingface_hub>=0.32.0"
+  changed_when: false
+
+- name: Run llama server model sync from Hugging Face
+  ansible.windows.win_command:
+    cmd: >-
+      "{{ __llama_venv_python }}" "{{ __llama_sync_script }}"
+      --models-file "{{ __llama_models_json }}"
+      --managed-dir "{{ __llama_managed_dir }}"
+      --links-dir "{{ __llama_links_dir }}"
+      --manifest-file "{{ __llama_manifest }}"
+      --preset-file "{{ __llama_preset_file }}"
+      --preset-global-file "{{ __llama_preset_global_json }}"
+      --container-links-dir "{{ __llama_links_dir | replace('\\', '/') }}"
+      {% if llama_server_models_prune_enabled | default(true) %}--prune{% endif %}
+      {% if llama_server_models_dry_run | default(false) %}--dry-run{% endif %}
+  register: __llama_sync_output__
+  changed_when: false
+  environment:
+    HF_TOKEN: "{{ lookup('env', 'HF_TOKEN') | default('', true) }}"
+
+- name: Parse llama server sync output
+  ansible.builtin.set_fact:
+    __llama_sync_result__: "{{ __llama_sync_output__.stdout | from_json }}"
+
+- name: Stat llama router preset file (when not in dry-run mode)
+  ansible.windows.win_stat:
+    path: "{{ __llama_preset_file }}"
+  register: __llama_preset_stat__
+  when: not (llama_server_models_dry_run | default(false))
+
+- name: Ensure router preset file exists when not in dry-run mode
+  ansible.builtin.assert:
+    that:
+      - __llama_preset_stat__.stat.exists
+    fail_msg: "llama router preset file was not created"
+  when: not (llama_server_models_dry_run | default(false))
+
+- name: Remove existing llama-server service if present
+  ansible.windows.win_service:
+    name: llama-server
+    state: absent
+  failed_when: false
+
+- name: Install llama-server service via NSSM
+  ansible.windows.win_command:
+    argv: >-
+      {{
+        [
+          __nssm_exe,
+          'install',
+          'llama-server',
+          (llama_server_windows_install_dir ~ '\\' ~ llama_server_windows_bin_name),
+          '--port',
+          (llama_server_port | default(11434) | string),
+          '--host',
+          (llama_server_host | default('127.0.0.1') | string),
+          '--metrics',
+          '--models-preset',
+          (__llama_preset_file | string),
+          '--models-max',
+          (llama_server_models_max | string)
+        ]
+        + (llama_server_argv_extra | default([]))
+      }}
+  args:
+    chdir: "{{ llama_server_windows_install_dir }}"
+  register: __llama_nssm_install
+  changed_when: __llama_nssm_install.rc == 0
+
+- name: Set llama-server AppDirectory via NSSM
+  ansible.windows.win_command:
+    argv:
+      - "{{ __nssm_exe }}"
+      - set
+      - llama-server
+      - AppDirectory
+      - "{{ llama_server_windows_install_dir }}"
+
+- name: Configure llama-server stdout log via NSSM
+  ansible.windows.win_command:
+    argv:
+      - "{{ __nssm_exe }}"
+      - set
+      - llama-server
+      - AppStdout
+      - "{{ llama_server_windows_install_dir }}\\llama-server.out.log"
+
+- name: Configure llama-server stderr log via NSSM
+  ansible.windows.win_command:
+    argv:
+      - "{{ __nssm_exe }}"
+      - set
+      - llama-server
+      - AppStderr
+      - "{{ llama_server_windows_install_dir }}\\llama-server.err.log"
+
+- name: Set llama-server Start mode to AUTO via NSSM
+  ansible.windows.win_command:
+    argv:
+      - "{{ __nssm_exe }}"
+      - set
+      - llama-server
+      - Start
+      - SERVICE_AUTO_START
+
+- name: Ensure llama-server service is running
+  ansible.windows.win_service:
+    name: llama-server
+    start_mode: auto
+    state: restarted
+  when: not ansible_check_mode | default(false)
+
+- name: Wait for llama-server to listen on port
+  ansible.windows.win_wait_for:
+    port: "{{ llama_server_port | default(11434) }}"
+    delay: 5
+    timeout: 60
+  when: not ansible_check_mode | default(false)
+
+- name: Query llama router model list
+  ansible.windows.win_uri:
+    url: "http://127.0.0.1:{{ llama_server_port | default(11434) }}/models?reload=1"
+    method: GET
+    status_code: 200
+    return_content: true
+  register: __llama_router_models__
+  changed_when: false
+  when:
+    - llama_server_models_sync_enabled | default(true)
+    - not (llama_server_models_dry_run | default(false))
+
+- name: Parse llama router model list payload
+  ansible.builtin.set_fact:
+    __llama_router_models_payload__: "{{ __llama_router_models__.content | from_json }}"
+  when:
+    - llama_server_models_sync_enabled | default(true)
+    - not (llama_server_models_dry_run | default(false))
+
+- name: Validate expected model aliases are available in router
+  ansible.builtin.assert:
+    that:
+      - item.name in (__llama_router_models_payload__.data | map(attribute='id') | list)
+    fail_msg: "Model name '{{ item.name }}' not present in llama router /models output"
+  loop: "{{ __llama_sync_result__.models | default([]) }}"
+  loop_control:
+    label: "{{ item.name }}"
+  when:
+    - llama_server_models_sync_enabled | default(true)
+    - not (llama_server_models_dry_run | default(false))
+
+- name: Set firewall remote IPs when string provided
+  ansible.builtin.set_fact:
+    llama_server_windows_remoteip: "{{ llama_server_windows_sources }}"
+  when:
+    - llama_server_windows_sources is defined
+    - llama_server_windows_sources is string
+
+- name: Set firewall remote IPs when list provided
+  ansible.builtin.set_fact:
+    llama_server_windows_remoteip: "{{ llama_server_windows_sources | join(',') }}"
+  when:
+    - llama_server_windows_sources is defined
+    - not (llama_server_windows_sources is string)
+
+- name: Allow incoming traffic to llama-server port through Windows Firewall
+  community.windows.win_firewall_rule:
+    name: Llama Server "{{ llama_server_port | default(11434) }}"
+    localport: "{{ llama_server_port | default(11434) }}"
+    protocol: tcp
+    direction: in
+    action: allow
+    enabled: true
+    # Provide one or more IPs/CIDRs in llama_server_windows_sources (e.g. ['203.0.113.10'])
+    remoteip: "{{ llama_server_windows_remoteip | default(omit) }}"
+    state: present
+
+# --- Windows Exporter tasks ---
+
+- name: Run Windows Exporter setup
+  when: llama_server_windows_exporter_enabled | default(true) | bool
+  ansible.builtin.include_tasks:
+    file: llama_exporter.yml
+  tags:
+    - roles::llama_server::llama_exporter
diff --git a/ansible/roles/windows_prometheus/tasks/main.yml b/ansible/roles/windows_prometheus/tasks/main.yml
index 88c6f30..b0ae8a5 100644
--- a/ansible/roles/windows_prometheus/tasks/main.yml
+++ b/ansible/roles/windows_prometheus/tasks/main.yml
@@ -110,7 +110,7 @@
   ansible.windows.win_service:
     name: Prometheus
     state: absent
-  ignore_errors: true
+  failed_when: false
 
 - name: Install Prometheus service via NSSM
   ansible.windows.win_command: >-
diff --git a/inventory/hosts b/inventory/hosts
index 0cce085..1388fc9 100644
--- a/inventory/hosts
+++ b/inventory/hosts
@@ -121,6 +121,7 @@ steambox.local
 
 [llamacpp]
 gpu-01.lab.alexpires.me
+ci-01.lab.alexpires.me
 
 [qwen3_6_host]
 gpu-01.lab.alexpires.me