Setup (Windows 11, Intel Arc Pro B60 24GB)

The scripts assume this layout, so keep the paths or update them in governor/profiles.json, llama-swap/config.yaml and the wrappers:

C:\AI\llama.cpp\vulkan-intelfa\   engine build (engine\build.ps1)
C:\AI\llama.cpp\vulkan\           optional: official llama.cpp Vulkan release (qwen engine stock)
C:\AI\models\qwen3.8-27b\         Qwen3.8-27B-UD-Q4_K_M.gguf, mmproj-F16.gguf, Qwen3.8-27B-DFlash2-Q4_K_M.gguf
C:\AI\llama-swap\                 llama-swap.exe, config.yaml, api-key.txt
C:\AI\governor\                   everything from governor\ (+ voice\voice.html in governor\public\)
C:\AI\bench\                      everything from bench\ (benchmarks write their logs here)
C:\AI\evalsuite\                  everything from evalsuite\
C:\AI\tts\  C:\AI\whisper\         created by voice\setup-*.ps1
C:\AI\claude-local\               agents\claude-local\ (LOCAL-SYSTEM.md, mcp.json)
%USERPROFILE%\bin\                agents\claude-qwen.cmd, agents\codex-qwen.cmd

1. Engine

powershell -ExecutionPolicy Bypass -File engine\build.ps1

This needs git and Python on PATH. It downloads a portable toolchain, checks and applies the two vendored Intel FA patches (engine/patches/) to llama.cpp b11524, and writes C:\AI\llama.cpp\vulkan-intelfa\llama-server.exe. Our own build log shows 3 min 21 s from start to BUILD OK on a run where the toolchain was already in place (build.log); a first run also downloads the toolchain. On Linux use engine/build-linux.sh (patch application verified; not run on Linux with an Arc GPU by the author). Background: INTEL-FA.md.

2. Models

Download Qwen3.8-27B-UD-Q4_K_M.gguf and mmproj-F16.gguf (vision) from unsloth/Qwen3.8-27B-GGUF on Hugging Face, plus the DFlash2 draft GGUF (Qwen3.8-27B-DFlash2-Q4_K_M.gguf), into C:\AI\models\qwen3.8-27b\. We did not record where we downloaded the drafter from, so look for a Qwen3.8-27B DFlash2 draft on Hugging Face and check its licence; if you know the right source, please tell us. Optional extras: Qwen3.8-27B-Q4_0.gguf for the turbo variant, and an MTP draft for the 262K variant.

3. Key and server

$b = New-Object byte[] 24; [Security.Cryptography.RandomNumberGenerator]::Create().GetBytes($b); $k = ([BitConverter]::ToString($b) -replace '-').ToLower()
setx LOCAL_QWEN_KEY $k; $k | Set-Content C:\AI\llama-swap\api-key.txt

Put llama-swap.exe and llama-swap/config.yaml in C:\AI\llama-swap\. Then register two logon tasks, both started through conhost.exe --headless so no console windows pop up:

conhost.exe --headless "C:\AI\llama-swap\llama-swap.exe" --config "C:\AI\llama-swap\config.yaml" --listen 127.0.0.1:8090
conhost.exe --headless powershell.exe -NoProfile -NonInteractive -ExecutionPolicy Bypass -File "C:\AI\governor\governor.ps1"

4. Verify

qwen doctor      # 26 checks: server, governor, VRAM headroom, no-context-shift, voice, Tailscale routes, tasks, agents, power, disks, last benchmark
qwen bench       # 3-turn agent test at 30-37K context. We measured 55-57 s, coding 7/7, tools 3/3 (details: INTEL-FA.md)
qwen snapshot "first working config"

qwen is governor\qwen.ps1. Add a qwen.cmd shim on PATH that calls it.

5. Remote (optional): Tailscale

tailscale serve --bg --https=8445 http://127.0.0.1:8090                       # API for other PCs, tailnet only
tailscale serve --bg --https=8445 --set-path=/governor/ C:\AI\governor\public  # status page + voice page

Keep everything tailnet-only. Don't use tailscale funnel for the model API.

5b. A chat UI, voice and tools (optional)

PLATFORM.md adds Open WebUI on top of this setup, with voice in and out (VOICE.md), MCP tools and phone access over Tailscale, and the settings that keep it from disturbing the prompt cache.

6. Nightly regression (optional)

Schedule governor\nightly.ps1 daily at 3 am. It runs the doctor, the benchmark and a 14-task agent sample, and writes governor\public\nightly.json. It skips itself while an evaluation is running or the model is busy.