diff --git a/llm-server/llm/ansible/README.md b/llm-server/llm/ansible/README.md
new file mode 100644
index 0000000..b49908b
--- /dev/null
+++ b/llm-server/llm/ansible/README.md
@@ -0,0 +1,54 @@
+# LLM serving stack – Ansible control project
+
+Executes the LLM serving stack for exactly one environment on
+`ghostnet-ki` (GNAI-008 section 7.4). Host baseline, Podman runtime,
+workload accounts and AMD GPU access are managed by
+[`../../host/ansible`](../../host/ansible/README.md); this project never
+changes host state.
+
+## Layout
+
+- `inventories/<environment>/` – one inventory per KI environment
+  (currently `testing`); deployment uses the `ansible-deploy` account with
+  `become: true`.
+- `playbooks/ai.yml` – applies `model_store` then `inference_engine`.
+- `roles/model_store/` – shared, pinned model weights under
+  `/var/lib/ghostnet-ai/shared/models` (GNAI-010 sections 3 and 4).
+- `roles/inference_engine/` – rootless Quadlets, network, image provisioning
+  and the exclusive profile switch for llama.cpp (GNAI-010 sections 5 and 6).
+
+## Control node setup
+
+Use the repository-wide control venv built from
+[`../../../shared/ansible/requirements-control.txt`](../../../shared/ansible/requirements-control.txt)
+(`ansible-core==2.21.4`), for example:
+
+```bash
+python3 -m venv .venv
+.venv/bin/pip install -r shared/ansible/requirements-control.txt
+cd llm-server/llm/ansible
+../../../.venv/bin/ansible-playbook -i inventories/testing playbooks/ai.yml --syntax-check
+```
+
+## Usage (Testing)
+
+```bash
+# Preview only – no changes, no downloads, no service start:
+../../../.venv/bin/ansible-playbook -i inventories/testing playbooks/ai.yml --check --diff
+
+# Provision model weights only (tag models):
+../../../.venv/bin/ansible-playbook -i inventories/testing playbooks/ai.yml --tags models
+
+# Render quadlets and start/verify the selected profile (tag service):
+../../../.venv/bin/ansible-playbook -i inventories/testing playbooks/ai.yml --tags service
+```
+
+`inference_engine_active_profile` selects which rendered quadlet is started;
+all other profiles stay rendered but stopped. Check mode downloads no model
+and starts no service. A second identical run must report `changed=0`
+(GNAI-010 section 10.4).
+
+The full technical acceptance (64k/128k start, VRAM distribution, prefill
+ladder, restart behaviour) is an operator-run protocol documented in
+[GNAI-010](../../../docs/platform/llm-server/ai/GNAI-010-TECH_Llama-Serving.md);
+this role automates deployment and basic health only.
diff --git a/llm-server/llm/ansible/ansible.cfg b/llm-server/llm/ansible/ansible.cfg
new file mode 100644
index 0000000..b13fcf0
--- /dev/null
+++ b/llm-server/llm/ansible/ansible.cfg
@@ -0,0 +1,12 @@
+[defaults]
+home = .ansible
+host_key_checking = True
+interpreter_python = auto_silent
+local_tmp = .ansible/tmp
+retry_files_enabled = False
+roles_path = roles:../../../shared/ansible/roles
+timeout = 15
+
+[ssh_connection]
+pipelining = True
+ssh_args = -o ControlMaster=auto -o ControlPersist=60s
diff --git a/llm-server/llm/ansible/inventories/testing/group_vars/llm_servers.yml b/llm-server/llm/ansible/inventories/testing/group_vars/llm_servers.yml
new file mode 100644
index 0000000..dd5641c
--- /dev/null
+++ b/llm-server/llm/ansible/inventories/testing/group_vars/llm_servers.yml
@@ -0,0 +1,14 @@
+---
+# Environment-specific values for the Testing stack (GNAI-008 section 7.4,
+# GNAI-010 section 5). The workload account, quadlet directory, unit prefix
+# and network exist once per environment; Production and Experimental use the
+# same role logic with their own values.
+inference_engine_user: gnai-testing
+inference_engine_uid: 981
+inference_engine_user_home: /var/lib/ghostnet-ai/testing
+inference_engine_unit_prefix: gnai-testing-llama
+inference_engine_network_name: gnai-testing-llama
+
+# GNAI-008 section 7.5: bring-up starts at 64k; the 128k target profile is
+# accepted separately (GNAI-010 section 10.3). Set to null after acceptance.
+inference_engine_ctx_size_override: 65536
diff --git a/llm-server/llm/ansible/inventories/testing/hosts.yml b/llm-server/llm/ansible/inventories/testing/hosts.yml
new file mode 100644
index 0000000..8c61b48
--- /dev/null
+++ b/llm-server/llm/ansible/inventories/testing/hosts.yml
@@ -0,0 +1,7 @@
+all:
+  children:
+    llm_servers:
+      hosts:
+        ghostnet-ki:
+          ansible_host: 10.0.0.90
+          ansible_user: ansible-deploy
diff --git a/llm-server/llm/ansible/playbooks/ai.yml b/llm-server/llm/ansible/playbooks/ai.yml
new file mode 100644
index 0000000..79983d8
--- /dev/null
+++ b/llm-server/llm/ansible/playbooks/ai.yml
@@ -0,0 +1,16 @@
+---
+# Deploys the LLM serving stack for exactly one explicitly selected
+# environment (GNAI-008 section 7.4). Host baseline, Podman runtime,
+# workload accounts and GPU access are managed by the host project
+# (llm-server/host/ansible); this playbook never changes host state.
+#
+# Usage (Testing):
+#   ansible-playbook -i inventories/testing playbooks/ai.yml --check --diff
+#   ansible-playbook -i inventories/testing playbooks/ai.yml --tags models
+#   ansible-playbook -i inventories/testing playbooks/ai.yml --tags service
+- name: Bring the LLM serving stack of one environment into its desired state
+  hosts: llm_servers
+  become: true
+  roles:
+    - role: model_store
+    - role: inference_engine
diff --git a/llm-server/llm/ansible/roles/inference_engine/README.md b/llm-server/llm/ansible/roles/inference_engine/README.md
new file mode 100644
index 0000000..430267c
--- /dev/null
+++ b/llm-server/llm/ansible/roles/inference_engine/README.md
@@ -0,0 +1,54 @@
+# inference_engine
+
+Deploys the inference engine for one KI environment as rootless Podman
+Quadlets (GNAI-010 sections 5 and 6; first profile: llama.cpp). Runs after
+`model_store` in the same play and reuses `model_store_root`.
+
+## Variables
+
+- `inference_engine_image_ref` – immutable image digest (`ghcr.io/...@sha256:...`).
+- `inference_engine_user` / `inference_engine_uid` / `inference_engine_user_home`
+  – workload account created by the host `workload_accounts` role.
+- `inference_engine_unit_prefix` – unit and container prefix of the
+  environment (Testing: `gnai-testing-llama`).
+- `inference_engine_network_name` / `inference_engine_network_alias` –
+  internal-only Quadlet network (`Internal=true`), alias `llama-cpp` for the
+  later API policy layer.
+- `inference_engine_profiles` – map of model profiles; each needs `unit_suffix`,
+  `model_directory`, `model_file`, `alias`, `ctx_size`, `n_predict`,
+  `reasoning_effort`, sampling values and an `mtp` block.
+- `inference_engine_active_profile` – the single profile to start.
+- `inference_engine_ctx_size_override` – when not `null`, replaces every
+  profile's `ctx_size` (Testing bring-up: 65536; set to `null` after the 128k
+  acceptance, GNAI-008 section 7.5).
+
+## Behaviour
+
+- Validates declarations, requires `/dev/kfd` plus exactly two render nodes,
+  and requires every profile's model artifact (run `--tags models` first).
+- Pulls the pinned image by digest into the environment's rootless store
+  only when missing.
+- Renders `<network>.network` and one `.container` per profile into
+  `/etc/containers/systemd/users/<uid>/`; units are never enabled for
+  `default.target`.
+- Units mutually exclude each other via `Conflicts=`/`Before=`
+  (GNAI-010 section 5.3).
+- Exercises the Quadlet generator with `--dryrun` before any reload; quadlet
+  changes notify a `daemon-reload` of the workload user manager. Services
+  are never restarted automatically.
+- Switch/start (only with the `service` tag, never in check mode): stop any
+  other active engine unit, start the selected unit, wait for `/health`, and
+  verify the served alias via `/v1/models`. The HTTP check uses whichever of
+  `curl`/`wget` exists in the pinned image (GNAI-010 section 8).
+
+## Tags
+
+- `service` – everything in this role. Model provisioning belongs to the
+  `models` tag of the `model_store` role.
+
+## Notes
+
+- There is no `PublishPort=`; the engine is reachable only inside the
+  environment network.
+- The role does not start the engine in check mode and never downloads
+  models.
diff --git a/llm-server/llm/ansible/roles/inference_engine/defaults/main.yml b/llm-server/llm/ansible/roles/inference_engine/defaults/main.yml
new file mode 100644
index 0000000..9fffd9e
--- /dev/null
+++ b/llm-server/llm/ansible/roles/inference_engine/defaults/main.yml
@@ -0,0 +1,88 @@
+---
+# Inference engine deployment (GNAI-008 section 7.4, GNAI-010 sections 3, 5,
+# 6 and 9). The first engine profile is llama.cpp; Strata follows as an
+# Experimental candidate and is intentionally not modelled here yet.
+#
+# Container image: immutable digest, never a moving tag (GNAI-010 section 3.1).
+inference_engine_image: ghcr.io/ggml-org/llama.cpp:server-rocm
+inference_engine_image_ref: ghcr.io/ggml-org/llama.cpp@sha256:b429b8be10d3d0353a09c98f4ab840fe516f779858a28f9fd4e5bdf10bc63cb4
+inference_engine_image_build: "b11058"
+inference_engine_image_revision: f072b103714dfa1eee531f80b24512faf38e3dd2
+
+# Environment identity (overridden per inventory, GNAI-008 section 7.4).
+inference_engine_user: gnai-testing
+inference_engine_uid: 981
+inference_engine_user_home: /var/lib/ghostnet-ai/testing
+inference_engine_unit_prefix: gnai-testing-llama
+inference_engine_network_name: gnai-testing-llama
+inference_engine_network_alias: llama-cpp
+
+# Paths
+inference_engine_quadlet_user_dir: /etc/containers/systemd/users
+inference_engine_generator_path: /usr/lib/systemd/system-generators/podman-system-generator
+inference_engine_container_port: 8080
+
+# Bring-up context override: when set, it replaces the profile ctx_size for
+# every profile (GNAI-008 section 7.5: 64k bring-up before the 128k target).
+# Set to null once the 128k acceptance (GNAI-010 section 10.3) has passed.
+inference_engine_ctx_size_override: null
+
+# Shared serving parameters (GNAI-010 sections 6.1-6.5, 8).
+inference_engine_common:
+  n_gpu_layers: all
+  split_mode: layer
+  tensor_split: "1,1"
+  flash_attn: "on"
+  cache_type_k: q8_0
+  cache_type_v: q8_0
+  reasoning: "on"
+  reasoning_budget: -1
+
+# Exactly one profile is active at a time; the other quadlet stays rendered
+# but stopped (GNAI-010 section 5.3). Unit names follow GNAI-010 section 5.1.
+inference_engine_active_profile: swift-qwen3.8-27b-q6k
+
+inference_engine_profiles:
+  swift-qwen3.8-27b-q6k:
+    unit_suffix: swift
+    description: Swift efficiency finetune of Qwen3.8-27B (Q6_K)
+    model_directory: swift-qwen3.8-27b
+    model_file: Swift-Qwen3.8-27B-Q6_K.gguf
+    alias: swift-qwen3.8-27b-q6k
+    ctx_size: 131072
+    n_predict: 32768
+    parallel: 1
+    reasoning_effort: xhigh
+    temp: 1.0
+    top_p: 0.95
+    top_k: 20
+    min_p: 0.0
+    presence_penalty: 0.0
+    repeat_penalty: 1.0
+    mtp:
+      enabled: false
+      spec_draft_n_max: 2
+  unsloth-qwen3.8-27b-ud-q6k:
+    unit_suffix: unsloth
+    description: Unsloth UD quant of Qwen3.8-27B (Q6_K)
+    model_directory: unsloth-qwen3.8-27b
+    model_file: Qwen3.8-27B-UD-Q6_K.gguf
+    alias: unsloth-qwen3.8-27b-ud-q6k
+    ctx_size: 131072
+    n_predict: 32768
+    parallel: 1
+    # xhigh for the controlled A/B comparison; switch to medium for the
+    # practice comparison (GNAI-010 section 6.1).
+    reasoning_effort: xhigh
+    temp: 1.0
+    top_p: 0.95
+    top_k: 20
+    min_p: 0.0
+    presence_penalty: 0.0
+    repeat_penalty: 1.0
+    mtp:
+      enabled: false
+      spec_draft_n_max: 2
+
+# Derived paths (do not override).
+inference_engine_quadlet_dir: "{{ inference_engine_quadlet_user_dir }}/{{ inference_engine_uid }}"
diff --git a/llm-server/llm/ansible/roles/inference_engine/handlers/main.yml b/llm-server/llm/ansible/roles/inference_engine/handlers/main.yml
new file mode 100644
index 0000000..0201ef8
--- /dev/null
+++ b/llm-server/llm/ansible/roles/inference_engine/handlers/main.yml
@@ -0,0 +1,20 @@
+---
+# Reloads the persistent user manager of the workload account after quadlet
+# changes (GNAI-010 section 9). Handlers never run in check mode. Services
+# are intentionally not restarted automatically: switching profiles is an
+# explicit sequential operation (GNAI-010 section 5.3).
+- name: Reload the workload user manager
+  ansible.builtin.command:
+    argv:
+      - /usr/sbin/runuser
+      - --user
+      - "{{ inference_engine_user }}"
+      - --
+      - /usr/bin/env
+      - "HOME={{ inference_engine_user_home }}"
+      - "XDG_RUNTIME_DIR=/run/user/{{ inference_engine_uid }}"
+      - /usr/bin/systemctl
+      - --user
+      - daemon-reload
+    chdir: "{{ inference_engine_user_home }}"
+  changed_when: false
diff --git a/llm-server/llm/ansible/roles/inference_engine/tasks/main.yml b/llm-server/llm/ansible/roles/inference_engine/tasks/main.yml
new file mode 100644
index 0000000..43c8532
--- /dev/null
+++ b/llm-server/llm/ansible/roles/inference_engine/tasks/main.yml
@@ -0,0 +1,388 @@
+---
+# Deploys the inference engine (llama.cpp profile) for one environment
+# (GNAI-010 sections 5, 6 and 9). Rendering and provisioning run under the
+# `service` tag, model artifacts under the `models` tag of the model_store
+# role. The selected unit is only started by the explicit switch/start tasks
+# below and is never enabled for default.target (GNAI-010 section 5.3).
+- name: Validate inference engine declarations
+  ansible.builtin.assert:
+    that:
+      - inference_engine_user | length > 0
+      - inference_engine_uid | int > 0
+      - inference_engine_user_home is match("^/")
+      - inference_engine_unit_prefix | length > 0
+      - inference_engine_network_name | length > 0
+      - inference_engine_image_ref is match("^ghcr\\.io/.+@sha256:[0-9a-f]{64}$")
+      - inference_engine_image_revision is match("^[0-9a-f]{40}$")
+      - inference_engine_profiles | length >= 2
+      - inference_engine_active_profile in inference_engine_profiles
+      - model_store_root is match("^/")
+    fail_msg: Inference engine declarations are incomplete or ambiguous.
+  tags:
+    - service
+
+- name: Derive the systemd unit name of every profile
+  ansible.builtin.set_fact:
+    inference_engine_unit_names: >-
+      {{ inference_engine_unit_names | default({})
+         | combine({ item.key:
+                      inference_engine_unit_prefix ~ '-' ~ item.value.unit_suffix }) }}
+  loop: "{{ inference_engine_profiles | dict2items }}"
+  loop_control:
+    label: "{{ item.key }}"
+  tags:
+    - service
+
+- name: Require the rootless runtime utilities
+  ansible.builtin.stat:
+    path: "{{ item }}"
+  loop:
+    - /usr/sbin/runuser
+    - /usr/bin/podman
+    - /usr/bin/systemctl
+    - "{{ inference_engine_generator_path }}"
+  register: inference_engine_tools
+  tags:
+    - service
+
+- name: Verify the rootless runtime utilities are executable
+  ansible.builtin.assert:
+    that:
+      - item.stat.exists
+      - item.stat.executable
+    loop: "{{ inference_engine_tools.results }}"
+    fail_msg: >-
+      {{ item.item }} is missing or not executable; run the host baseline
+      roles first (llm-server/host/ansible).
+  tags:
+    - service
+
+- name: Require the KFD compute device
+  ansible.builtin.stat:
+    path: /dev/kfd
+  register: inference_engine_kfd
+  tags:
+    - service
+
+- name: Discover the AMD render nodes
+  ansible.builtin.find:
+    paths: /dev/dri
+    patterns: renderD*
+  register: inference_engine_render_node_find
+  tags:
+    - service
+
+- name: Record the render nodes
+  ansible.builtin.set_fact:
+    inference_engine_render_nodes: >-
+      {{ inference_engine_render_node_find.files
+         | map(attribute='path') | sort }}
+  tags:
+    - service
+
+- name: Verify KFD and exactly two render nodes
+  ansible.builtin.assert:
+    that:
+      - inference_engine_kfd.stat.exists
+      - inference_engine_render_nodes | length == 2
+    fail_msg: >-
+      Expected /dev/kfd and exactly two render nodes (2x RX 6900 XT
+      baseline); found {{ inference_engine_render_nodes | length }}
+      render node(s): {{ inference_engine_render_nodes }}.
+  tags:
+    - service
+
+- name: Require every profile model artifact to be present
+  ansible.builtin.stat:
+    path: "{{ model_store_root }}/{{ item.value.model_directory }}/{{ item.value.model_file }}"
+  loop: "{{ inference_engine_profiles | dict2items }}"
+  loop_control:
+    label: "{{ item.key }}"
+  register: inference_engine_models
+  tags:
+    - service
+
+- name: Verify every profile model artifact is present
+  ansible.builtin.assert:
+    that:
+      - item.stat.exists
+    loop: "{{ inference_engine_models.results }}"
+    fail_msg: >-
+      Model artifact for {{ item.item.key }} is missing under
+      {{ model_store_root }}; run the model_store role (--tags models) first.
+  tags:
+    - service
+
+- name: Check whether the pinned engine image is in the workload store
+  ansible.builtin.command:
+    argv:
+      - /usr/sbin/runuser
+      - --user
+      - "{{ inference_engine_user }}"
+      - --
+      - /usr/bin/env
+      - "HOME={{ inference_engine_user_home }}"
+      - "XDG_RUNTIME_DIR=/run/user/{{ inference_engine_uid }}"
+      - /usr/bin/podman
+      - image
+      - exists
+      - "{{ inference_engine_image_ref }}"
+    chdir: "{{ inference_engine_user_home }}"
+  changed_when: false
+  failed_when: inference_engine_image_present.rc not in [0, 1]
+  check_mode: false
+  register: inference_engine_image_present
+  tags:
+    - service
+
+- name: Pull the pinned engine image into the workload store
+  ansible.builtin.command:
+    argv:
+      - /usr/sbin/runuser
+      - --user
+      - "{{ inference_engine_user }}"
+      - --
+      - /usr/bin/env
+      - "HOME={{ inference_engine_user_home }}"
+      - "XDG_RUNTIME_DIR=/run/user/{{ inference_engine_uid }}"
+      - /usr/bin/podman
+      - pull
+      - "{{ inference_engine_image_ref }}"
+    chdir: "{{ inference_engine_user_home }}"
+  when:
+    - not ansible_check_mode
+    - inference_engine_image_present.rc != 0
+  tags:
+    - service
+
+- name: Ensure the environment quadlet directory exists
+  ansible.builtin.file:
+    path: "{{ inference_engine_quadlet_dir }}"
+    state: directory
+    owner: root
+    group: root
+    mode: "0755"
+  tags:
+    - service
+
+- name: Render the environment Podman network quadlet
+  ansible.builtin.template:
+    src: llama-cpp.network.j2
+    dest: "{{ inference_engine_quadlet_dir }}/{{ inference_engine_network_name }}.network"
+    owner: root
+    group: root
+    mode: "0644"
+  notify: Reload the workload user manager
+  tags:
+    - service
+
+- name: Render one engine quadlet per model profile
+  ansible.builtin.template:
+    src: llama-server.container.j2
+    dest: "{{ inference_engine_quadlet_dir }}/{{ inference_engine_unit_names[quadlet_item.key] }}.container"
+    owner: root
+    group: root
+    mode: "0644"
+  vars:
+    entry:
+      profile_key: "{{ quadlet_item.key }}"
+      profile: "{{ quadlet_item.value }}"
+      unit_name: "{{ inference_engine_unit_names[quadlet_item.key] }}"
+      container_name: "{{ inference_engine_unit_names[quadlet_item.key] }}"
+      ctx_size: >-
+        {{ (inference_engine_ctx_size_override | int)
+           if inference_engine_ctx_size_override is not none
+           else quadlet_item.value.ctx_size }}
+      other_units: >-
+        {{ inference_engine_unit_names.values()
+           | difference([inference_engine_unit_names[quadlet_item.key]]) }}
+  loop: "{{ inference_engine_profiles | dict2items }}"
+  loop_control:
+    loop_var: quadlet_item
+    label: "{{ quadlet_item.key }}"
+  notify: Reload the workload user manager
+  tags:
+    - service
+
+- name: Exercise the Quadlet generator without starting anything
+  ansible.builtin.command:
+    argv:
+      - /usr/sbin/runuser
+      - --user
+      - "{{ inference_engine_user }}"
+      - --
+      - /usr/bin/env
+      - "HOME={{ inference_engine_user_home }}"
+      - "XDG_RUNTIME_DIR=/run/user/{{ inference_engine_uid }}"
+      - "QUADLET_UNIT_DIRS={{ inference_engine_quadlet_dir }}"
+      - "{{ inference_engine_generator_path }}"
+      - --dryrun
+    chdir: "{{ inference_engine_user_home }}"
+  changed_when: false
+  when: not ansible_check_mode
+  tags:
+    - service
+
+- name: Read the current state of every engine unit
+  ansible.builtin.command:
+    argv:
+      - /usr/sbin/runuser
+      - --user
+      - "{{ inference_engine_user }}"
+      - --
+      - /usr/bin/env
+      - "HOME={{ inference_engine_user_home }}"
+      - "XDG_RUNTIME_DIR=/run/user/{{ inference_engine_uid }}"
+      - /usr/bin/systemctl
+      - --user
+      - is-active
+      - "{{ inference_engine_unit_names[item.key] }}"
+    chdir: "{{ inference_engine_user_home }}"
+  changed_when: false
+  failed_when: inference_engine_unit_states.rc not in [0, 3]
+  check_mode: false
+  loop: "{{ inference_engine_profiles | dict2items }}"
+  loop_control:
+    label: "{{ item.key }}"
+  register: inference_engine_unit_states
+  when: not ansible_check_mode
+  tags:
+    - service
+
+- name: Stop any other active engine unit before the switch
+  ansible.builtin.command:
+    argv:
+      - /usr/sbin/runuser
+      - --user
+      - "{{ inference_engine_user }}"
+      - --
+      - /usr/bin/env
+      - "HOME={{ inference_engine_user_home }}"
+      - "XDG_RUNTIME_DIR=/run/user/{{ inference_engine_uid }}"
+      - /usr/bin/systemctl
+      - --user
+      - stop
+      - "{{ inference_engine_unit_names[item.item.key] }}"
+    chdir: "{{ inference_engine_user_home }}"
+  when:
+    - not ansible_check_mode
+    - item.stdout | trim != 'active'
+    - item.item.key != inference_engine_active_profile
+  loop: "{{ inference_engine_unit_states.results }}"
+  loop_control:
+    label: "{{ item.item.key }}"
+  tags:
+    - service
+
+- name: Start the selected engine unit
+  ansible.builtin.command:
+    argv:
+      - /usr/sbin/runuser
+      - --user
+      - "{{ inference_engine_user }}"
+      - --
+      - /usr/bin/env
+      - "HOME={{ inference_engine_user_home }}"
+      - "XDG_RUNTIME_DIR=/run/user/{{ inference_engine_uid }}"
+      - /usr/bin/systemctl
+      - --user
+      - start
+      - "{{ inference_engine_unit_names[inference_engine_active_profile] }}"
+    chdir: "{{ inference_engine_user_home }}"
+  when:
+    - not ansible_check_mode
+    - >-
+      (inference_engine_unit_states.results
+       | selectattr('item.key', 'equalto', inference_engine_active_profile)
+       | map(attribute='stdout') | join('') | trim) != 'active'
+  tags:
+    - service
+
+- name: Detect an HTTP client inside the engine container
+  ansible.builtin.command:
+    argv:
+      - /usr/sbin/runuser
+      - --user
+      - "{{ inference_engine_user }}"
+      - --
+      - /usr/bin/env
+      - "HOME={{ inference_engine_user_home }}"
+      - "XDG_RUNTIME_DIR=/run/user/{{ inference_engine_uid }}"
+      - /usr/bin/podman
+      - exec
+      - "{{ inference_engine_unit_names[inference_engine_active_profile] }}"
+      - /bin/sh
+      - -c
+      - "command -v curl || command -v wget || true"
+    chdir: "{{ inference_engine_user_home }}"
+  changed_when: false
+  check_mode: false
+  register: inference_engine_http_client
+  when: not ansible_check_mode
+  tags:
+    - service
+
+- name: Require an HTTP client in the pinned engine image
+  ansible.builtin.assert:
+    that:
+      - inference_engine_http_client.stdout | trim in ['curl', 'wget']
+    fail_msg: >-
+      The pinned engine image contains neither curl nor wget; health checks
+      cannot run (GNAI-010 section 8 requires checking the image tool first).
+  when: not ansible_check_mode
+  tags:
+    - service
+
+- name: Wait for the engine health endpoint
+  ansible.builtin.command:
+    argv: "{{ [ '/usr/sbin/runuser', '--user', inference_engine_user, '--',
+               '/usr/bin/env', 'HOME=' ~ inference_engine_user_home,
+               'XDG_RUNTIME_DIR=/run/user/' ~ inference_engine_uid,
+               '/usr/bin/podman', 'exec',
+               inference_engine_unit_names[inference_engine_active_profile],
+               '/bin/sh', '-c' ] +
+             (['curl', '-fsS', health_url] if inference_engine_http_client.stdout | trim == 'curl'
+              else ['wget', '-qO-', health_url]) }}"
+  vars:
+    health_url: "http://127.0.0.1:{{ inference_engine_container_port }}/health"
+  changed_when: false
+  check_mode: false
+  register: inference_engine_health
+  until: inference_engine_health.rc == 0
+  retries: 60
+  delay: 5
+  when: not ansible_check_mode
+  tags:
+    - service
+
+- name: Verify the served model alias
+  ansible.builtin.command:
+    argv: "{{ [ '/usr/sbin/runuser', '--user', inference_engine_user, '--',
+               '/usr/bin/env', 'HOME=' ~ inference_engine_user_home,
+               'XDG_RUNTIME_DIR=/run/user/' ~ inference_engine_uid,
+               '/usr/bin/podman', 'exec',
+               inference_engine_unit_names[inference_engine_active_profile],
+               '/bin/sh', '-c' ] +
+             (['curl', '-fsS', models_url] if inference_engine_http_client.stdout | trim == 'curl'
+              else ['wget', '-qO-', models_url]) }}"
+  vars:
+    models_url: "http://127.0.0.1:{{ inference_engine_container_port }}/v1/models"
+  changed_when: false
+  check_mode: false
+  register: inference_engine_models_endpoint
+  when: not ansible_check_mode
+  tags:
+    - service
+
+- name: Verify the active profile alias is served
+  ansible.builtin.assert:
+    that:
+      - >-
+        inference_engine_profiles[inference_engine_active_profile].alias
+        in inference_engine_models_endpoint.stdout
+    fail_msg: >-
+      The /v1/models response does not contain the expected alias
+      {{ inference_engine_profiles[inference_engine_active_profile].alias }}.
+  when: not ansible_check_mode
+  tags:
+    - service
diff --git a/llm-server/llm/ansible/roles/inference_engine/templates/llama-cpp.network.j2 b/llm-server/llm/ansible/roles/inference_engine/templates/llama-cpp.network.j2
new file mode 100644
index 0000000..0860320
--- /dev/null
+++ b/llm-server/llm/ansible/roles/inference_engine/templates/llama-cpp.network.j2
@@ -0,0 +1,4 @@
+[Network]
+NetworkName={{ inference_engine_network_name }}
+Driver=bridge
+Internal=true
diff --git a/llm-server/llm/ansible/roles/inference_engine/templates/llama-server.container.j2 b/llm-server/llm/ansible/roles/inference_engine/templates/llama-server.container.j2
new file mode 100644
index 0000000..38cf4e2
--- /dev/null
+++ b/llm-server/llm/ansible/roles/inference_engine/templates/llama-server.container.j2
@@ -0,0 +1,27 @@
+[Unit]
+Description=LLM inference engine ({{ entry.profile_key }})
+{% for other in entry.other_units %}
+Conflicts={{ other }}
+{% endfor %}
+{% if entry.other_units | length == 1 %}
+Before={{ entry.other_units | first }}
+{% endif %}
+
+[Container]
+Image={{ inference_engine_image_ref }}
+Pull=never
+ContainerName={{ entry.container_name }}
+AddDevice=/dev/kfd
+{% for device in inference_engine_render_nodes %}
+AddDevice={{ device }}
+{% endfor %}
+GroupAdd=keep-groups
+Volume={{ model_store_root }}/{{ entry.profile.model_directory }}:/models:ro
+Network={{ inference_engine_network_name }}
+NetworkAlias={{ inference_engine_network_alias }}
+Exec=llama-server --host 0.0.0.0 --port {{ inference_engine_container_port }} --model /models/{{ entry.profile.model_file }} --alias {{ entry.profile.alias }} --parallel {{ entry.profile.parallel }} --ctx-size {{ entry.ctx_size }} --n-predict {{ entry.profile.n_predict }} --n-gpu-layers {{ inference_engine_common.n_gpu_layers }} --split-mode {{ inference_engine_common.split_mode }} --tensor-split {{ inference_engine_common.tensor_split }} --flash-attn {{ inference_engine_common.flash_attn }} --cache-type-k {{ inference_engine_common.cache_type_k }} --cache-type-v {{ inference_engine_common.cache_type_v }} --jinja --reasoning {{ inference_engine_common.reasoning }} --reasoning-effort {{ entry.profile.reasoning_effort }} --reasoning-budget {{ inference_engine_common.reasoning_budget }} --temp {{ entry.profile.temp }} --top-p {{ entry.profile.top_p }} --top-k {{ entry.profile.top_k }} --min-p {{ entry.profile.min_p }} --presence-penalty {{ entry.profile.presence_penalty }} --repeat-penalty {{ entry.profile.repeat_penalty }}{% if entry.profile.mtp.enabled %} --spec-type draft-mtp --spec-draft-n-max {{ entry.profile.mtp.spec_draft_n_max }}{% endif %} --no-webui --metrics
+
+[Service]
+Restart=on-failure
+RestartSec=10
+TimeoutStartSec=900
diff --git a/llm-server/llm/ansible/roles/model_store/README.md b/llm-server/llm/ansible/roles/model_store/README.md
new file mode 100644
index 0000000..196c141
--- /dev/null
+++ b/llm-server/llm/ansible/roles/model_store/README.md
@@ -0,0 +1,35 @@
+# model_store
+
+Provides the shared, pinned model store under
+`/var/lib/ghostnet-ai/shared/models` (GNAI-010 sections 3 and 4). Weights are
+large, immutable, non-secret artifacts; every KI environment mounts its
+profile directory read-only.
+
+## Variables
+
+- `model_store_root` – root of the shared store (default
+  `/var/lib/ghostnet-ai/shared/models`).
+- `model_store_artifacts` – list of pinned artifacts with `profile`,
+  `directory`, `file`, `sha256`, `url`, `repository`, `revision`, `origin`
+  and `license`.
+- `model_store_include_vision` – additionally provisions the Swift vision
+  projector when `true` (default `false`; vision stage only, GNAI-010
+  section 3.4).
+- `model_store_required_free_mib` – free-space reserve checked before any
+  pending download (default 71680).
+
+## Behaviour
+
+- Manifest is validated (unique profiles, 64-hex checksums, HTTPS URLs).
+- Directories are `root:root 0755`, artifact files `root:root 0644`; a
+  `SOURCE.yml` provenance record is rendered per directory.
+- Downloads go to a hidden `.part` file with `get_url` checksum verification
+  and are moved into place only after the checksum passed (atomic).
+- An existing file with a mismatching checksum fails the run; it is never
+  overwritten or used.
+- Check mode downloads nothing; already-present files are still verified.
+
+## Tags
+
+`models` – everything in this role. The `service` tag of `inference_engine`
+also re-verifies present artifacts.
diff --git a/llm-server/llm/ansible/roles/model_store/defaults/main.yml b/llm-server/llm/ansible/roles/model_store/defaults/main.yml
new file mode 100644
index 0000000..c24b563
--- /dev/null
+++ b/llm-server/llm/ansible/roles/model_store/defaults/main.yml
@@ -0,0 +1,51 @@
+---
+# Shared, pinned model store (GNAI-010 sections 3 and 4). Weights are large,
+# immutable, non-secret artifacts; every AI environment mounts them read-only.
+model_store_root: /var/lib/ghostnet-ai/shared/models
+
+# Minimum free space on the target filesystem before downloading missing
+# artifacts (GNAI-010 section 4: ~45 GiB final stock, the largest single
+# temporary download and an operating buffer).
+model_store_required_free_mib: 71680
+
+# The vision projector is intentionally not part of the first text/tool
+# bring-up (GNAI-010 section 3.4). Enable only for the dedicated vision stage.
+model_store_include_vision: false
+
+model_store_artifacts:
+  - profile: swift-qwen3.8-27b-q6k
+    directory: swift-qwen3.8-27b
+    file: Swift-Qwen3.8-27B-Q6_K.gguf
+    sha256: 7f4de8abd5446c08b0f975a1b38e43d02639f4ae9ecdd2ddfd5c5b0f612bda59
+    url: https://huggingface.co/ukisai/Swift-Qwen3.8-27B-GGUF/resolve/eb0e3a7dc70643c2ab920f5133750d02c20849ff/Swift-Qwen3.8-27B-Q6_K.gguf
+    repository: ukisai/Swift-Qwen3.8-27B-GGUF
+    revision: eb0e3a7dc70643c2ab920f5133750d02c20849ff
+    origin: Swift efficiency finetune of Qwen3.8-27B (UkisAI), Q6_K
+    license: >-
+      Qwen base model Apache-2.0; Swift changes under Swift Open License 1.0
+      (free personal, research, education, evaluation and commercial use up to
+      the revenue limit defined in the model card).
+  - profile: unsloth-qwen3.8-27b-ud-q6k
+    directory: unsloth-qwen3.8-27b
+    file: Qwen3.8-27B-UD-Q6_K.gguf
+    sha256: c9c206812fbe4ac7b76a729e25928b63f2ae89d37f69da7a71c20aec763cd436
+    url: https://huggingface.co/unsloth/Qwen3.8-27B-GGUF/resolve/4ca720788d1e01f1bff70c033e0d0028fd02e502/Qwen3.8-27B-UD-Q6_K.gguf
+    repository: unsloth/Qwen3.8-27B-GGUF
+    revision: 4ca720788d1e01f1bff70c033e0d0028fd02e502
+    origin: Unsloth dynamic quant (UD) of Qwen3.8-27B, Q6_K
+    license: Apache-2.0
+
+model_store_vision_artifact:
+  profile: swift-qwen3.8-27b-vision-mmproj
+  directory: swift-qwen3.8-27b
+  file: mmproj-Swift-Qwen3.8-27B-F16.gguf
+  sha256: daa1116c9422fa390cc8688495da0e91781f92841dfc3b31a378ff252571745a
+  url: https://huggingface.co/ukisai/Swift-Qwen3.8-27B-GGUF/resolve/eb0e3a7dc70643c2ab920f5133750d02c20849ff/mmproj-Swift-Qwen3.8-27B-F16.gguf
+  repository: ukisai/Swift-Qwen3.8-27B-GGUF
+  revision: eb0e3a7dc70643c2ab920f5133750d02c20849ff
+  origin: Swift vision projector, F16 (vision stage only)
+  license: >-
+    Qwen base model Apache-2.0; Swift changes under Swift Open License 1.0.
+
+model_store_directory_mode: "0755"
+model_store_file_mode: "0644"
diff --git a/llm-server/llm/ansible/roles/model_store/tasks/main.yml b/llm-server/llm/ansible/roles/model_store/tasks/main.yml
new file mode 100644
index 0000000..f906682
--- /dev/null
+++ b/llm-server/llm/ansible/roles/model_store/tasks/main.yml
@@ -0,0 +1,176 @@
+---
+# Builds the shared, pinned model store (GNAI-010 sections 3 and 4).
+# Downloads are atomic and checksum-verified; an existing file with a
+# mismatching checksum aborts the run instead of being overwritten.
+- name: Assemble the effective artifact manifest
+  ansible.builtin.set_fact:
+    model_store_all_artifacts: >-
+      {{ model_store_artifacts
+          + (model_store_vision_artifact | ternary(
+              [model_store_vision_artifact], []))
+          if model_store_include_vision
+          else model_store_artifacts }}
+  tags:
+    - models
+    - service
+
+- name: Validate the artifact manifest
+  ansible.builtin.assert:
+    that:
+      - model_store_root is match("^/")
+      - model_store_all_artifacts | length > 0
+      - model_store_all_artifacts | map(attribute='profile') | list | unique | length
+        == model_store_all_artifacts | length
+      - model_store_all_artifacts | map(attribute='directory') | list | length > 0
+      - model_store_all_artifacts | map(attribute='file') | list | length > 0
+      - model_store_all_artifacts | map(attribute='sha256') | select('match', '^[0-9a-f]{64}$') | list | length
+        == model_store_all_artifacts | length
+      - model_store_all_artifacts | map(attribute='url') | select('match', '^https://') | list | length
+        == model_store_all_artifacts | length
+    fail_msg: The model store manifest is incomplete or ambiguous.
+  tags:
+    - models
+    - service
+
+- name: Require the checksum utility
+  ansible.builtin.stat:
+    path: /usr/bin/sha256sum
+  register: model_store_sha256sum_tool
+  tags:
+    - models
+    - service
+
+- name: Verify the checksum utility is executable
+  ansible.builtin.assert:
+    that:
+      - model_store_sha256sum_tool.stat.exists
+      - model_store_sha256sum_tool.stat.executable
+    fail_msg: /usr/bin/sha256sum is required to verify model artifacts.
+  tags:
+    - models
+    - service
+
+- name: Create the model profile directories
+  ansible.builtin.file:
+    path: "{{ model_store_root }}/{{ item }}"
+    state: directory
+    owner: root
+    group: root
+    mode: "{{ model_store_directory_mode }}"
+  loop: "{{ model_store_all_artifacts | map(attribute='directory') | unique }}"
+  tags:
+    - models
+    - service
+
+- name: Check which model artifacts are already present
+  ansible.builtin.stat:
+    path: "{{ model_store_root }}/{{ item.directory }}/{{ item.file }}"
+    get_checksum: false
+  loop: "{{ model_store_all_artifacts }}"
+  loop_control:
+    label: "{{ item.profile }}"
+  register: model_store_present
+  tags:
+    - models
+    - service
+
+- name: Require sufficient free space before downloading missing artifacts
+  ansible.builtin.command:
+    argv:
+      - /bin/df
+      - -m
+      - --output=avail
+      - "{{ model_store_root }}"
+  changed_when: false
+  check_mode: false
+  register: model_store_free_space
+  when: >-
+    model_store_present.results
+    | selectattr('stat.exists', 'equalto', false)
+    | list | length > 0
+  tags:
+    - models
+
+- name: Verify the free-space reserve for pending downloads
+  ansible.builtin.assert:
+    that:
+      - model_store_free_space.stdout_lines | last | int >= model_store_required_free_mib | int
+    fail_msg: >-
+      Less than {{ model_store_required_free_mib }} MiB free on
+      {{ model_store_root }}; refusing to download model artifacts
+      (GNAI-010 section 4).
+  when: model_store_free_space is not skipped
+  tags:
+    - models
+
+- name: Verify the checksum of already present artifacts
+  ansible.builtin.command:
+    argv:
+      - /usr/bin/sha256sum
+      - "{{ model_store_root }}/{{ item.item.directory }}/{{ item.item.file }}"
+  changed_when: false
+  check_mode: false
+  loop: "{{ model_store_present.results | selectattr('stat.exists', 'equalto', true) | list }}"
+  loop_control:
+    label: "{{ item.item.profile }}"
+  register: model_store_existing_checksums
+  failed_when: model_store_existing_checksums.stdout.split()[0] != item.item.sha256
+  tags:
+    - models
+    - service
+
+- name: Download missing model artifacts atomically
+  ansible.builtin.get_url:
+    url: "{{ item.item.url }}"
+    dest: "{{ model_store_root }}/{{ item.item.directory }}/.{{ item.item.file }}.part"
+    checksum: "sha256:{{ item.item.sha256 }}"
+    mode: "0644"
+    owner: root
+    group: root
+    timeout: 120
+  loop: "{{ model_store_present.results | selectattr('stat.exists', 'equalto', false) | list }}"
+  loop_control:
+    label: "{{ item.item.profile }}"
+  when: not ansible_check_mode
+  register: model_store_downloads
+  tags:
+    - models
+
+- name: Move verified downloads to their final model path
+  ansible.builtin.command:
+    argv:
+      - /bin/mv
+      - "{{ model_store_root }}/{{ item.item.directory }}/.{{ item.item.file }}.part"
+      - "{{ model_store_root }}/{{ item.item.directory }}/{{ item.item.file }}"
+  loop: "{{ model_store_present.results | selectattr('stat.exists', 'equalto', false) | list }}"
+  loop_control:
+    label: "{{ item.item.profile }}"
+  when: not ansible_check_mode
+  tags:
+    - models
+
+- name: Enforce model artifact ownership and permissions
+  ansible.builtin.file:
+    path: "{{ model_store_root }}/{{ item.directory }}/{{ item.file }}"
+    state: file
+    owner: root
+    group: root
+    mode: "{{ model_store_file_mode }}"
+  loop: "{{ model_store_all_artifacts }}"
+  loop_control:
+    label: "{{ item.profile }}"
+  tags:
+    - models
+    - service
+
+- name: Write the per-directory origin manifest SOURCE.yml
+  ansible.builtin.template:
+    src: source.yml.j2
+    dest: "{{ model_store_root }}/{{ item }}/SOURCE.yml"
+    owner: root
+    group: root
+    mode: "{{ model_store_file_mode }}"
+  loop: "{{ model_store_all_artifacts | map(attribute='directory') | unique }}"
+  tags:
+    - models
+    - service
diff --git a/llm-server/llm/ansible/roles/model_store/templates/source.yml.j2 b/llm-server/llm/ansible/roles/model_store/templates/source.yml.j2
new file mode 100644
index 0000000..7084c2c
--- /dev/null
+++ b/llm-server/llm/ansible/roles/model_store/templates/source.yml.j2
@@ -0,0 +1,15 @@
+# Provenance record for the artifacts of this model directory.
+# Rendered by the model_store role; do not edit on the host.
+{% set entries = model_store_all_artifacts | selectattr('directory', 'equalto', item) | list %}
+{% for entry in entries %}
+{{ entry.profile }}:
+  file: {{ entry.file }}
+  sha256: {{ entry.sha256 }}
+  repository: {{ entry.repository }}
+  revision: {{ entry.revision }}
+  url: {{ entry.url }}
+  origin: >-
+    {{ entry.origin }}
+  license: >-
+    {{ entry.license | replace("\n", " ") }}
+{% endfor %}