diff --git a/infrastructure/playbooks/windows_ollama_models.yml b/infrastructure/playbooks/windows_ollama_models.yml
new file mode 100644
index 0000000..669d12e
--- /dev/null
+++ b/infrastructure/playbooks/windows_ollama_models.yml
@@ -0,0 +1,162 @@
+---
+# windows_ollama_models.yml — Manage local Ollama model variants on Windows GPU hosts.
+#
+# This is the small, Ansible-like control surface for the two Windows NVIDIA
+# machines that run Ollama for Prole embedding workloads.
+#
+# Usage:
+#   cd infrastructure
+#   ansible-playbook playbooks/windows_ollama_models.yml --ask-pass
+#
+# Single host:
+#   ansible-playbook playbooks/windows_ollama_models.yml --limit morgoth.prole.org --ask-pass
+#
+# Dry-run intent only:
+#   ansible-playbook playbooks/windows_ollama_models.yml --check --diff --ask-pass
+#
+# Inventory target:
+#   [workstations_windows]
+#   morgoth.prole.org
+#   fairyland.prole.org
+#
+# What this deploys:
+#   ollama create mxbai-embed-large-2k FROM mxbai-embed-large PARAMETER num_ctx 2048
+#
+# The desired Modelfile is written to C:\ProgramData\Prole\Ollama\Models and a
+# SHA-256 marker is recorded after successful creation. Re-running the playbook
+# is safe: it recreates the model only when the target model is missing or the
+# managed Modelfile content changes.
+
+- name: Configure Prole Ollama models on Windows GPU hosts
+  hosts: workstations_windows
+  gather_facts: true
+
+  vars:
+    prole_ollama_models_dir: 'C:\ProgramData\Prole\Ollama\Models'
+    prole_ollama_models:
+      - name: mxbai-embed-large-2k
+        base: mxbai-embed-large
+        parameters:
+          num_ctx: 2048
+
+    # Override per-host if WinRM cannot see ollama.exe in PATH.
+    # Common installs include:
+    #   C:\Program Files\Ollama\ollama.exe
+    #   C:\Users\<user>\AppData\Local\Programs\Ollama\ollama.exe
+    ollama_windows_exe: ''
+
+  tasks:
+    - name: Resolve ollama.exe path
+      ansible.windows.win_shell: |
+        $ErrorActionPreference = 'Stop'
+        $candidates = @()
+
+        if ('{{ ollama_windows_exe }}') {
+          $candidates += '{{ ollama_windows_exe }}'
+        }
+
+        $cmd = Get-Command ollama.exe -ErrorAction SilentlyContinue
+        if ($cmd) {
+          $candidates += $cmd.Source
+        }
+
+        $candidates += @(
+          'C:\Program Files\Ollama\ollama.exe',
+          'C:\Users\{{ ansible_user }}\AppData\Local\Programs\Ollama\ollama.exe'
+        )
+
+        foreach ($candidate in $candidates | Select-Object -Unique) {
+          if ($candidate -and (Test-Path -LiteralPath $candidate)) {
+            Write-Output $candidate
+            exit 0
+          }
+        }
+
+        throw 'ollama.exe was not found. Set ollama_windows_exe for this host or install Ollama first.'
+      args:
+        executable: powershell.exe
+      register: ollama_exe_result
+      changed_when: false
+
+    - name: Set resolved Ollama executable fact
+      ansible.builtin.set_fact:
+        ollama_resolved_exe: "{{ ollama_exe_result.stdout_lines[0] }}"
+
+    - name: Ensure Prole Ollama model state directory exists
+      ansible.windows.win_file:
+        path: "{{ prole_ollama_models_dir }}"
+        state: directory
+
+    - name: Show planned Ollama model configuration in check mode
+      ansible.builtin.debug:
+        msg: >-
+          Would ensure {{ item.name }} exists from {{ item.base }} with num_ctx={{ item.parameters.num_ctx }}
+          using {{ ollama_resolved_exe }}.
+      loop: "{{ prole_ollama_models }}"
+      when: ansible_check_mode
+
+    - name: Create or update managed Ollama model variants
+      ansible.windows.win_shell: |
+        $ErrorActionPreference = 'Stop'
+
+        $ollama = '{{ ollama_resolved_exe }}'
+        $modelName = '{{ item.name }}'
+        $baseModel = '{{ item.base }}'
+        $modelDir = '{{ prole_ollama_models_dir }}'
+        $modelfilePath = Join-Path $modelDir "$modelName.Modelfile"
+        $markerPath = Join-Path $modelDir "$modelName.Modelfile.sha256"
+
+        $modelfile = @"
+        FROM {{ item.base }}
+        PARAMETER num_ctx {{ item.parameters.num_ctx }}
+        "@.TrimStart()
+
+        Set-Content -LiteralPath $modelfilePath -Value $modelfile -Encoding ascii -NoNewline
+        $desiredHash = (Get-FileHash -Algorithm SHA256 -LiteralPath $modelfilePath).Hash
+        $recordedHash = if (Test-Path -LiteralPath $markerPath) {
+          (Get-Content -LiteralPath $markerPath -Raw).Trim()
+        } else {
+          ''
+        }
+
+        & $ollama show $modelName *> $null
+        $modelExists = ($LASTEXITCODE -eq 0)
+
+        if ($modelExists -and ($recordedHash -eq $desiredHash)) {
+          Write-Output "unchanged: $modelName already matches managed Modelfile"
+          exit 0
+        }
+
+        & $ollama pull $baseModel
+        if ($LASTEXITCODE -ne 0) {
+          throw "ollama pull failed for $baseModel"
+        }
+
+        & $ollama create $modelName -f $modelfilePath
+        if ($LASTEXITCODE -ne 0) {
+          throw "ollama create failed for $modelName"
+        }
+
+        Set-Content -LiteralPath $markerPath -Value $desiredHash -Encoding ascii -NoNewline
+        Write-Output "changed: created or updated $modelName from $baseModel"
+      args:
+        executable: powershell.exe
+      loop: "{{ prole_ollama_models }}"
+      register: ollama_model_results
+      changed_when: "'changed:' in ollama_model_results.stdout"
+      when: not ansible_check_mode
+
+    - name: Verify managed Ollama model variants
+      ansible.windows.win_shell: |
+        $ErrorActionPreference = 'Stop'
+        & '{{ ollama_resolved_exe }}' list
+      args:
+        executable: powershell.exe
+      register: ollama_list_result
+      changed_when: false
+      when: not ansible_check_mode
+
+    - name: Show Ollama model list
+      ansible.builtin.debug:
+        var: ollama_list_result.stdout_lines
+      when: not ansible_check_mode