Recipe by Mia's AI Lab @MiaAI_lab · based on turboderp, Qwen, DeepSeek, z-lab, r0b0tlab

Original recipe · Site · GitHub · X

License: Kit scripts: MIT · turboderp: MIT (engine), Apache-2.0 (quants) · Qwen: Apache-2.0 · DeepSeek: MIT · z-lab: MIT.

TESTED BY Mia's AI Lab on one RTX 5090 (32 GB), Linux, driver 610.43.02, ExLlamaV3 1.6.0 (pinned) · their results (benchmark table)

Qwen3.8-27B one-click kit for NVIDIA cards

EASY on Windows: one double-click. Linux: 2 commands.

FORNVIDIA card, 16 to 32 GB. Windows or Linux.

NEED64-bit Python 3.11 or newer.

NEEDNode 22.19 or newer, for the chat app.

NEEDNVIDIA driver 570 or newer.

GETDownload the kit from GitHub: Code > Download ZIP. Unzip it.

WINDOWSDouble-click windows\START-HERE.bat

LINUX

bash linux/setup.sh
bash linux/start.sh

OPENThe chat opens in your browser. Use the full address it prints, with the token.

ON 24 GBIt picks 4.0 bpw, 262K context, images on.

NEXT TIMEWindows: windows\start.bat. Linux: bash linux/start.sh

HER TUNED SETTINGS FOR A 24 GB CARD

QUANT4.0 bpw (pick it in setup; it is the 24 GB default)

FILE.env (in the kit folder)

CONTEXT_SIZE=262144  # full 262K context
GPU_MEM_GB=22        # room for driver
CACHE_QUANT=4        # int4 KV, near-lossless
DRAFT=mtp            # faster replies
HOST=127.0.0.1       # keeps API private

Setup writes these for you when you take the 24 GB default. Paste only to set them by hand. From README 1, README 2, README 3.

SAFETYThe API is open to your network by default. On Wi-Fi you don't own, set HOST=127.0.0.1 in .env.

SAFETYThe chat app's agent can run commands on this PC. Pick a workspace folder before the first message.

OR · LET A CODING AGENT DO IT

PASTEInto Claude Code, Codex or Cursor:

Set up the Qwen3.8-27B one-click kit by Mia's AI Lab on this computer: https://github.com/MiaAI-Lab/Qwen3.8-27B-16gb-NVIDIA-GPUs-one-click-install
Follow its README exactly.
1. Check first and tell me what you find: NVIDIA card with 16 to 32 GB of VRAM, driver 570+, 64-bit Python 3.11+, Node 22.19+. Stop if any is missing.
2. Do not install anything system-wide or as admin. The kit keeps everything in its own folder.
3. Run the kit's setup, then start it. Accept the size it recommends.
4. In .env, make sure these are set: CONTEXT_SIZE=262144, GPU_MEM_GB=22, CACHE_QUANT=4, DRAFT=mtp, HOST=127.0.0.1 (keep HOST=127.0.0.1 unless I say this network is mine).
5. Run the kit's doctor command and show me the result.
6. Give me the full chat address it prints, token included.
Ask me before deleting anything.

If it fails

It loads but replies crawl (Windows).
Close what holds graphics memory (browser, games, Discord, other AI apps). The start script lists them. Or lower CONTEXT_SIZE in .env.
"Images: off" when it starts.
The image part did not fit. Lower CONTEXT_SIZE and restart, or pick a smaller size.
Linux: "Insufficient VRAM in split for model and cache".
Lower CONTEXT_SIZE or GPU_MEM_GB in .env, or run ./linux/simplex setup and pick a smaller size.
The chat says "authentication required".
Open the full address printed in the window or in logs/, with the token.
Anything else.
Windows: windows\simplex.bat doctor. Linux: ./linux/simplex doctor

From the original README, shortened. Read it for full details.

What they tested

Speed (RigMark 1.3.0, median of 5 runs) and quality (MMLU 200 questions, GSM8K 100) on the RTX 5090 · evidence
How much context fits, measured under 14.7 GB and 22.1 GB memory caps, Sep 6, 2026. The README does not name the card for these runs. · evidence
16 GB baseline: memory use measured at load under a 14.7 GB cap · evidence
Engine version pinned (ExLlamaV3 v1.6.0); start scripts refuse other versions; unit tests in tests/ · evidence

Not shown in their results: No published Windows test results. No published run on a 24 GB card; the 24 GB default comes from the 22.1 GB cap measurement. No CI.

RunCards does not test recipes. We only list recipes whose authors publish their own test results. Checked 2026-10-19.

Who made what

turboderp: ExLlamaV3 engine and the EXL3 quants (2.5 bpw and up) (also)
Qwen: the Qwen3.8-27B model
DeepSeek: DeepSeek Harness, the chat app it opens
z-lab: DFlash2 (optional speed-up, 32 GB cards only)
r0b0tlab: EXL3 quant of the DFlash2 drafter (optional)

The kit's README names these in its intro, its Engine line and its Credits line.