Qwen3.8-27B on Windows 11 + Arc Pro B70: the 19 August upgrade

If you already followed Windows 11 hosts — Qwen3.8-27B you are done with the hard part. Docker Desktop is installed. The 18 GB model is on disk. The server answers at http://127.0.0.1:8000/v1.

This is the next image, not a new kit.

The kits were devised and end-to-end tested by Ian Hudson (aitesthive.com) on 18 August. The 19 August overlay is the same digest, same checkpoint, three extra patches, and prefix cache on for real sessions.

Cookbook:

• Windows standalone guide
• Qwen3.8 vLLM XPU recipe
• Model: SergiioB/Qwen3.8-27B-GPTQ-Int4-sym-G128-MTP-BF16

What you get

	18 August kit	19 August image
Model files	already on disk	reuse them
Display-safe VRAM	0.75 + 4.25 GiB fp8 KV, 100K	unchanged
Draft	BF16 MTP head as stored	same files; draft requantized to INT4 at start
Prefix cache	off	on (chat, tools, system prompt)
Linux C1 p512/g128 n=5	81.20 matched BF16-draft arm (83.7 on the published BF16 row)	112.65 draft-INT4, cache off

Do not copy 112.65 onto a Windows table. Ian’s Docker number was ~70 tok/s on the 18 August BF16-draft kit (self-report, one machine). Re-run .\Test-CookbookDecode.ps1 after the rebuild if you want a Windows number.

Prefix cache does not make the first unique prompt faster. It is for the turns after that. Leave it on for Open WebUI, Pi, anything with a system prompt. Turn it off only for a cold unique-prompt decode test.

Docker Desktop — the actual steps

Work in the kit folder. If you cloned the cookbook that is windows/Qwen38-Docker-Standalone/. If you unzipped Ian’s zip, pull or copy the updated scripts into that folder first.

Set-ExecutionPolicy -Scope Process Bypass
.\Upgrade-Qwen38-Docker.ps1

That is the whole upgrade:

1. Builds qwen38-b70-docker:2026.08.19 (five hash-checked patches).
2. Deletes the old qwen38-b70-docker-test container.
3. Starts a new one with draft-INT4 on and prefix cache on.
4. Leaves .\models\Qwen3.8-27B-GPTQ-Int4 alone.

Restarting the 18 August container is not an upgrade. Docker keeps the old entrypoint until you recreate.

Manual equivalent:

.\Build-Qwen38Image-Docker.ps1
.\Start-Qwen38-Docker.ps1 -Recreate

Confirm the overlay actually ran:

docker logs qwen38-b70-docker-test | findstr /C:"draft-INT4"

You want:

[start] draft-INT4 S+M1 overlay ENABLED

Cold decode test (prefix cache off, matches the Linux speed card):

.\Start-Qwen38-Docker.ps1 -Recreate -PrefixCache 0
.\Test-CookbookDecode.ps1

Then put prefix cache back on for daily use:

.\Start-Qwen38-Docker.ps1 -Recreate

Back to the 18 August draft (same model files):

.\Start-Qwen38-Docker.ps1 -Recreate -DraftInt4 0

Endpoint is still http://127.0.0.1:8000/v1, model name qwen38.

WSLC — only if that is the kit you run

WSLC stays the experimental path (~26 tok/s on 18 August). Same overlay, same recreate idea:

Set-ExecutionPolicy -Scope Process Bypass
.\Upgrade-Qwen38-WSLC.ps1

Confirm with wslc logs qwen38-b70-friendly. This overlay does not fix the WSLC slowdown.

Fresh machine (no 18 August kit)

Install Docker Desktop with the Linux/WSL 2 engine, then:

Set-ExecutionPolicy -Scope Process Bypass
.\Setup-Qwen38-Docker.ps1

Setup now builds 2026.08.19. You still wait through the first model download (~18.2 GiB) and the tea-length load from the Windows bind mount.

What not to change

• Do not raise GPU_MEMORY_UTILIZATION above 0.75 or drop the 4.25 GiB KV pin on a display-attached B70. Ian already measured the failure: auto-sized KV ate the desktop and decode fell to ~26 tok/s.
• Do not publish 112.65 as a Windows result.
• Do not overwrite the Linux BF16-draft LocalMaxxing row (83.7). Draft-INT4 is a second row.

Sampling follows the checkpoint (--generation-config auto): thinking 1.0 / 0.95 / 20; non-thinking 0.7 / 0.80 / 20 with presence penalty 1.5. That is what Qwen documents, not greedy.