Instructions to use unsloth/Qwen3.6-27B-MTP-GGUF with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use unsloth/Qwen3.6-27B-MTP-GGUF with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("image-text-to-text", model="unsloth/Qwen3.6-27B-MTP-GGUF")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
pipe(text=messages)

# Load model directly
from transformers import AutoModel
model = AutoModel.from_pretrained("unsloth/Qwen3.6-27B-MTP-GGUF", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

llama.cpp

How to use unsloth/Qwen3.6-27B-MTP-GGUF with llama.cpp:

Install (macOS, Linux)

curl -LsSf https://llama.app/install.sh | sh
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf unsloth/Qwen3.6-27B-MTP-GGUF:UD-Q4_K_XL
# Run inference directly in the terminal:
llama cli -hf unsloth/Qwen3.6-27B-MTP-GGUF:UD-Q4_K_XL

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf unsloth/Qwen3.6-27B-MTP-GGUF:UD-Q4_K_XL
# Run inference directly in the terminal:
llama cli -hf unsloth/Qwen3.6-27B-MTP-GGUF:UD-Q4_K_XL

Use pre-built binary

# Download pre-built binary from:
# https://github.com/ggerganov/llama.cpp/releases
# Start a local OpenAI-compatible server with a web UI:
./llama-server -hf unsloth/Qwen3.6-27B-MTP-GGUF:UD-Q4_K_XL
# Run inference directly in the terminal:
./llama-cli -hf unsloth/Qwen3.6-27B-MTP-GGUF:UD-Q4_K_XL

Build from source code

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-server llama-cli
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama-server -hf unsloth/Qwen3.6-27B-MTP-GGUF:UD-Q4_K_XL
# Run inference directly in the terminal:
./build/bin/llama-cli -hf unsloth/Qwen3.6-27B-MTP-GGUF:UD-Q4_K_XL

Use Docker

docker model run hf.co/unsloth/Qwen3.6-27B-MTP-GGUF:UD-Q4_K_XL

LM Studio
Jan

vLLM

How to use unsloth/Qwen3.6-27B-MTP-GGUF with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "unsloth/Qwen3.6-27B-MTP-GGUF"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "unsloth/Qwen3.6-27B-MTP-GGUF",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker

docker model run hf.co/unsloth/Qwen3.6-27B-MTP-GGUF:UD-Q4_K_XL

SGLang

How to use unsloth/Qwen3.6-27B-MTP-GGUF with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "unsloth/Qwen3.6-27B-MTP-GGUF" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "unsloth/Qwen3.6-27B-MTP-GGUF",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "unsloth/Qwen3.6-27B-MTP-GGUF" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "unsloth/Qwen3.6-27B-MTP-GGUF",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Ollama
How to use unsloth/Qwen3.6-27B-MTP-GGUF with Ollama:
```
ollama run hf.co/unsloth/Qwen3.6-27B-MTP-GGUF:UD-Q4_K_XL
```
Unsloth Desktop

How to use unsloth/Qwen3.6-27B-MTP-GGUF with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf unsloth/Qwen3.6-27B-MTP-GGUF:UD-Q4_K_XL

Configure the model in Pi

# Install Pi:
npm install -g @earendil-works/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "unsloth/Qwen3.6-27B-MTP-GGUF:UD-Q4_K_XL"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use unsloth/Qwen3.6-27B-MTP-GGUF with Docker Model Runner:
```
docker model run hf.co/unsloth/Qwen3.6-27B-MTP-GGUF:UD-Q4_K_XL
```

Lemonade

How to use unsloth/Qwen3.6-27B-MTP-GGUF with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull unsloth/Qwen3.6-27B-MTP-GGUF:UD-Q4_K_XL

Run and chat with the model

lemonade run user.Qwen3.6-27B-MTP-GGUF-UD-Q4_K_XL

List all available models

lemonade list

Hermes Agent

How to use unsloth/Qwen3.6-27B-MTP-GGUF with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf unsloth/Qwen3.6-27B-MTP-GGUF:UD-Q4_K_XL

Configure Hermes

# Install Hermes:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup
# Point Hermes at the local server:
hermes config set model.provider custom
hermes config set model.base_url http://127.0.0.1:8080/v1
hermes config set model.default unsloth/Qwen3.6-27B-MTP-GGUF:UD-Q4_K_XL

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use unsloth/Qwen3.6-27B-MTP-GGUF with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf unsloth/Qwen3.6-27B-MTP-GGUF:UD-Q4_K_XL

Configure OpenClaw

# Install OpenClaw:
npm install -g openclaw@latest
# Register the local server and set it as the default model:
openclaw onboard --non-interactive --mode local \
  --auth-choice custom-api-key \
  --custom-base-url http://127.0.0.1:8080/v1 \
  --custom-model-id "unsloth/Qwen3.6-27B-MTP-GGUF:UD-Q4_K_XL" \
  --custom-provider-id llama-cpp \
  --custom-compatibility openai \
  --custom-text-input \
  --accept-risk \
  --skip-health

Run OpenClaw

openclaw agent --local --agent main --message "Hello from Hugging Face"

Low VRAM<=8GB best result (40t/s WITHOUT MTP)

#17

by hellork - opened May 16

Discussion

hellork

May 16

•

edited May 20

Qwen 3.6 is already using every last bit of resources with 8GB cards.
If you have sufficient VRAM, please ignore. This is not for you.

The first time I tried this, no matter the settings, I was getting 2 tokens/sec. I must have done a bad build.
Now that --spec-type draft-mtp is part of the official llama.cpp, it appears to be working, sort of.

To get performance with full context is going to be a balancing act.
UD quants are the way to go. And 2- or 3-bit isn't really that bad.
I one-shot prompted two complete games, with sound and all, no errors, with 2-bit quants.

Currently, llama-server doesn't seem to honor different spec-type & spec-draft-n-max settings on a per-model basis in router mode.
So I ran the server with and without those options for these tests.

Hardware

0.00.244.412 I - CUDA0 : NVIDIA GeForce RTX 3070 Laptop GPU (7819 MiB, 29009 MiB free)
0.00.244.417 I - CPU : AMD Ryzen 9 5900HX with Radeon Graphics (31981 MiB, 31981 MiB free)

Default settings

[*]
jinja = true
kv-unified = true
temp = 0.0
top-p = 0.95
top-k = 20
min-p = 0.0
ctk = q4_0
ctv = q4_0

Results comparison

2-bit WITHOUT MTP

[Qwen3.6-35B-A3B-UD-Q2_K_XL]
reasoning-budget = 32000
c = 262144
ncmoe = 25

[51523] 0.40.479.772 I slot print_timing: id 0 | task 0 | prompt eval time = 5310.02 ms / 2189 tokens ( 2.43 ms per token, 412.24 tokens per second)
[51523] 0.40.479.778 I slot print_timing: id 0 | task 0 | eval time = 27769.75 ms / 1229 tokens ( 22.60 ms per token, 44.26 tokens per second)

(I got 24 tokens per second eval time before. It was because I had another llama.cpp process running in the background somehow.)

2-bit WITH MTP

Best result. (I had to bump up ncmoe to 30 to accommodate the extra MTP overhead. 7.7GB VRAM used)
[51530] 1.34.227.671 I slot print_timing: id 0 | task 0 | prompt eval time = 6498.15 ms / 2513 tokens ( 2.59 ms per token, 386.73 tokens per second)
[51530] 1.34.227.677 I slot print_timing: id 0 | task 0 | eval time = 77643.29 ms / 2566 tokens ( 30.26 ms per token, 33.05 tokens per second)

4-bit WITHOUT MTP

[Qwen3.6-35B-A3B-UD-Q4_K_XL]
reasoning-budget = 32000
c = 262144
ncmoe = 33

[63577] 1.51.410.621 I slot print_timing: id 0 | task 0 | prompt eval time = 18975.05 ms / 4383 tokens ( 4.33 ms per token, 230.99 tokens per second)
[63577] 1.51.410.627 I slot print_timing: id 0 | task 0 | eval time = 81967.32 ms / 3016 tokens ( 27.18 ms per token, 36.80 tokens per second)

3-bit WITH MTP

[Qwen3.6-35B-A3B-UD-IQ3_K_XL]
reasoning-budget = 32000
ctx-size = 262144
alias = qwen3.6q3
ncmoe = 35

Best result. (Again, the MTP overhead is not worth the cost. Here, the 3-bit model is out-performed by plain, vanilla 4-bit)
[64055] 1.52.818.880 I slot print_timing: id 0 | task 0 | prompt eval time = 5184.85 ms / 1013 tokens ( 5.12 ms per token, 195.38 tokens per second)
[64055] 1.52.818.887 I slot print_timing: id 0 | task 0 | eval time = 101676.40 ms / 3022 tokens ( 33.65 ms per token, 29.72 tokens per second)

llama-server --host 0.0.0.0 --port 8087 --models-dir mtp/ --models-preset mtp/.models.ini -np 1 --spec-type draft-mtp --spec-draft-n-max 6
--no-warmup --models-max 1 --sleep-idle-seconds 420 --jinja -t 6 --no-mmproj-offload

Conclusions

No speedup noticed with MTP on 8GB RTX 3070 mobile with full context window.
The extra gigabyte used by MTP means we need more -cpumoe to compensate.
This could be worked around by using a smaller context window. We get that.**
Earlier results with slow prompt eval were from having another llama.cpp process running in the background.
It's fairly-new. More optimization for low mem might be OTW.

Cheers!

hellork

May 16

•

edited May 16

Do not use -ngl 99. It [lllama.cpp] will auto-optimize without it. If you keep getting low t/s, try bumping up ncmoe instead.

szwedek

May 17

•

edited May 17

I've got rtx3070 and 400-700 prompt/25-32 gen tps using q5km mtp.
128k ctx

renatomoselli

May 17

I've got rtx3070 and 400-700 prompt/25-32 gen tps using q5km mtp.
128k ctx

please share your settings good sir

szwedek

May 17

please share your settings good sir

--host 0.0.0.0 -fa 1 -c 131072 --min-p 0.0 --fit 1 -b 2024 -ub 512 --no-mmap -ctk q8_0 -ctv q8_0 --jinja -m Qwen3.6-35B-A3B-UD-MTP-Q5_K_M.gguf --temp 0.7 --top-p 0.95 --top-k 20 --presence-penalty 0.0 --repeat-penalty 1.0 --chat-template-kwargs "{"preserve_thinking":true}" --no-mmproj -np 1 --spec-type draft-mtp --spec-draft-n-max 2 --chat-template-file qwen3.6_chat_template.txt --reasoning-budget 4096 --metrics --ctx-checkpoints 65 --checkpoint-every-n-tokens 4096

The template from here - https://huggingface.co/froggeric/Qwen-Fixed-Chat-Templates

renatomoselli

May 17

please share your settings good sir

--host 0.0.0.0 -fa 1 -c 131072 --min-p 0.0 --fit 1 -b 2024 -ub 512 --no-mmap -ctk q8_0 -ctv q8_0 --jinja -m Qwen3.6-35B-A3B-UD-MTP-Q5_K_M.gguf --temp 0.7 --top-p 0.95 --top-k 20 --presence-penalty 0.0 --repeat-penalty 1.0 --chat-template-kwargs "{"preserve_thinking":true}" --no-mmproj -np 1 --spec-type draft-mtp --spec-draft-n-max 2 --chat-template-file qwen3.6_chat_template.txt --reasoning-budget 4096 --metrics --ctx-checkpoints 65 --checkpoint-every-n-tokens 4096

The template from here - https://huggingface.co/froggeric/Qwen-Fixed-Chat-Templates

thanks, I just managed stable 35 t/s on two 3060 ti + 32gb ram with these specs, I'll play with your, as they seem better, and see what changes. I'll paste mine below

.\llama-server.exe ^
--models-preset "D:\Local_AI\custom_models\models.ini" ^
--models-dir "D:\Local_AI\custom_models" ^
--models-max 1 ^
--port 8080 ^
--host 127.0.0.2 ^
--fit on ^
--fit-target 128 ^
--mmap ^
--flash-attn on ^
--parallel 1 ^
--spec-type draft-mtp ^
--spec-draft-n-max 2 ^
--spec-draft-p-min 0.6 ^
--split-mode layer ^
--main-gpu 0 ^
--batch-size 512 ^
--ubatch-size 128

models.ini

[Qwen3.6-27B-MTP-UD-IQ3_XXS]
model=D:\Local_AI\custom_models\Qwen3.6-27B-MTP-UD-IQ3_XXS.gguf
jinja=true
ctx-size=81920
temperature=0.6
top-p=0.95
top-k=20
min-p=0.0
presence-penalty=0.0
repeat-penalty=1.0
cache-type-k=q4_0
cache-type-v=q4_0

hellork

May 19

Also, I read somewhere that draft token prediction is only good for stuff like coding, which has strict rules and surprisingly, many predictable tokens.
Creative writing, on the other hand, is free-flowing and very unpredictable. Supposedly won't benefit much from MTP, and could even be worse.

hellork changed discussion title from Don't Bother with VRAM<=8GB (+how to get 40t/s without it) to Low VRAM<=8GB challenge (+how to get 40t/s without MTP) May 19

hellork

May 20

--spec-draft-n-max 1 seems to be fastest in many cases.

szwedek

May 20

--spec-draft-n-max 1 seems to be fastest in many cases.

It depends on a quant.

hellork changed discussion title from Low VRAM<=8GB challenge (+how to get 40t/s without MTP) to Low VRAM<=8GB best result (40t/s WITHOUT MTP) May 20

Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.

Tap or paste here to upload images

· Sign up or log in to comment