Instructions to use ManniX-ITA/Qwen3.6-27B-A3B-Coder-MTP-GGUF with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use ManniX-ITA/Qwen3.6-27B-A3B-Coder-MTP-GGUF with Transformers:

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

pipe = pipeline("text-generation", model="ManniX-ITA/Qwen3.6-27B-A3B-Coder-MTP-GGUF")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoModel
model = AutoModel.from_pretrained("ManniX-ITA/Qwen3.6-27B-A3B-Coder-MTP-GGUF", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

llama.cpp

How to use ManniX-ITA/Qwen3.6-27B-A3B-Coder-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 ManniX-ITA/Qwen3.6-27B-A3B-Coder-MTP-GGUF:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf ManniX-ITA/Qwen3.6-27B-A3B-Coder-MTP-GGUF:Q4_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf ManniX-ITA/Qwen3.6-27B-A3B-Coder-MTP-GGUF:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf ManniX-ITA/Qwen3.6-27B-A3B-Coder-MTP-GGUF:Q4_K_M

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 ManniX-ITA/Qwen3.6-27B-A3B-Coder-MTP-GGUF:Q4_K_M
# Run inference directly in the terminal:
./llama-cli -hf ManniX-ITA/Qwen3.6-27B-A3B-Coder-MTP-GGUF:Q4_K_M

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 ManniX-ITA/Qwen3.6-27B-A3B-Coder-MTP-GGUF:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama-cli -hf ManniX-ITA/Qwen3.6-27B-A3B-Coder-MTP-GGUF:Q4_K_M

Use Docker

docker model run hf.co/ManniX-ITA/Qwen3.6-27B-A3B-Coder-MTP-GGUF:Q4_K_M

LM Studio
Jan

vLLM

How to use ManniX-ITA/Qwen3.6-27B-A3B-Coder-MTP-GGUF with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "ManniX-ITA/Qwen3.6-27B-A3B-Coder-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": "ManniX-ITA/Qwen3.6-27B-A3B-Coder-MTP-GGUF",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/ManniX-ITA/Qwen3.6-27B-A3B-Coder-MTP-GGUF:Q4_K_M

SGLang

How to use ManniX-ITA/Qwen3.6-27B-A3B-Coder-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 "ManniX-ITA/Qwen3.6-27B-A3B-Coder-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": "ManniX-ITA/Qwen3.6-27B-A3B-Coder-MTP-GGUF",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

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 "ManniX-ITA/Qwen3.6-27B-A3B-Coder-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": "ManniX-ITA/Qwen3.6-27B-A3B-Coder-MTP-GGUF",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Ollama
How to use ManniX-ITA/Qwen3.6-27B-A3B-Coder-MTP-GGUF with Ollama:
```
ollama run hf.co/ManniX-ITA/Qwen3.6-27B-A3B-Coder-MTP-GGUF:Q4_K_M
```
Unsloth Desktop

How to use ManniX-ITA/Qwen3.6-27B-A3B-Coder-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 ManniX-ITA/Qwen3.6-27B-A3B-Coder-MTP-GGUF:Q4_K_M

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": "ManniX-ITA/Qwen3.6-27B-A3B-Coder-MTP-GGUF:Q4_K_M"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use ManniX-ITA/Qwen3.6-27B-A3B-Coder-MTP-GGUF with Docker Model Runner:
```
docker model run hf.co/ManniX-ITA/Qwen3.6-27B-A3B-Coder-MTP-GGUF:Q4_K_M
```

Lemonade

How to use ManniX-ITA/Qwen3.6-27B-A3B-Coder-MTP-GGUF with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull ManniX-ITA/Qwen3.6-27B-A3B-Coder-MTP-GGUF:Q4_K_M

Run and chat with the model

lemonade run user.Qwen3.6-27B-A3B-Coder-MTP-GGUF-Q4_K_M

List all available models

lemonade list

Hermes Agent

How to use ManniX-ITA/Qwen3.6-27B-A3B-Coder-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 ManniX-ITA/Qwen3.6-27B-A3B-Coder-MTP-GGUF:Q4_K_M

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 ManniX-ITA/Qwen3.6-27B-A3B-Coder-MTP-GGUF:Q4_K_M

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use ManniX-ITA/Qwen3.6-27B-A3B-Coder-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 ManniX-ITA/Qwen3.6-27B-A3B-Coder-MTP-GGUF:Q4_K_M

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 "ManniX-ITA/Qwen3.6-27B-A3B-Coder-MTP-GGUF:Q4_K_M" \
  --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"

Best qwen finetune tested ever!

by repla73 - opened 17 days ago

Discussion

repla73

17 days ago

This model performed extremely well despite it's weight and even beat some fine-tuned quants of newer qwen3.8-27b on real-world agentic benchmark test (not imaginary numbers and charts, real tests).

So I spent a couple of days tuning ik_llama.cpp for CPU-only inference on a 4-core ARM Neoverse-N1.

Special thanks to Mannix-ITA team!
Results:
The final configuration improved trimmed token generation throughput by roughly:
+10.0 vs the previous MTP2 configuration
+9.0% vs the previous canonical MTP1 setup
Prompt processing stayed essentially unchanged.
The main win came from using an IQ4_KS copy of the output tensor for MTP speculative generation:
-mtprot iq4_ks
--spec-type mtp:n_max=2,p_min=0.0
This does not requantize the whole model. The normal output head remains Q6_K; ik_llama creates a separate ~4-bit MTP-only output head used for speculative token proposals.
Cost: about 259 MiB additional RAM.
***Compile configuration
The best build used Clang 20 with explicit Neoverse-N1 targeting:
-mcpu=neoverse-n1+dotprod+fp16
-O3
-ffast-math
-fno-finite-math-only
-flto
-pipe
Relevant build options:
GGML_NATIVE=OFF
GGML_OPENMP=ON
GGML_IQK_MUL_MAT=ON
GGML_IQK_FLASH_ATTENTION=ON
GGML_IQK_FA_ALL_QUANTS=ON
Explicit CPU targeting mattered more than relying on generic/native build assumptions.
***Best runtime configuration
context: 131072
batch: 2048
ubatch: 512
threads: 4
threads-batch: 4
flash-attn: on
K cache: q8_KV
V cache: q4_0
runtime repack: on
THP: on
reasoning: off
MTP p_min: 0.0
MTP depth: 2
MTP output: iq4_ks
The practical lesson: on ARM CPU inference, the biggest gains did not come from exotic kernel rewrites. They came from matching the build to the CPU, keeping all four physical cores busy, using the IQK path, tuning MTP depth, and reducing the cost of the speculative output projection.
After testing PGO, prefetching, unrolling and several Q4 kernel rewrites, most of them were neutral or slower. The simple runtime configuration above won.

There also seems to be useful memory headroom on CPU-only systems. I’d love to see whether a future MTP design could spend some of that on a stronger predictor/output head and convert the extra memory budget into higher draft acceptance and throughput.

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