Instructions to use unsloth/Kimi-K2.6-GGUF with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use unsloth/Kimi-K2.6-GGUF with Transformers:

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

pipe = pipeline("image-text-to-text", model="unsloth/Kimi-K2.6-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/Kimi-K2.6-GGUF", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

llama.cpp

How to use unsloth/Kimi-K2.6-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/Kimi-K2.6-GGUF:UD-Q4_K_XL
# Run inference directly in the terminal:
llama cli -hf unsloth/Kimi-K2.6-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/Kimi-K2.6-GGUF:UD-Q4_K_XL
# Run inference directly in the terminal:
llama cli -hf unsloth/Kimi-K2.6-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/Kimi-K2.6-GGUF:UD-Q4_K_XL
# Run inference directly in the terminal:
./llama-cli -hf unsloth/Kimi-K2.6-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/Kimi-K2.6-GGUF:UD-Q4_K_XL
# Run inference directly in the terminal:
./build/bin/llama-cli -hf unsloth/Kimi-K2.6-GGUF:UD-Q4_K_XL

Use Docker

docker model run hf.co/unsloth/Kimi-K2.6-GGUF:UD-Q4_K_XL

LM Studio
Jan

vLLM

How to use unsloth/Kimi-K2.6-GGUF with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "unsloth/Kimi-K2.6-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/Kimi-K2.6-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/Kimi-K2.6-GGUF:UD-Q4_K_XL

SGLang

How to use unsloth/Kimi-K2.6-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/Kimi-K2.6-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/Kimi-K2.6-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/Kimi-K2.6-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/Kimi-K2.6-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/Kimi-K2.6-GGUF with Ollama:
```
ollama run hf.co/unsloth/Kimi-K2.6-GGUF:UD-Q4_K_XL
```
Unsloth Desktop

How to use unsloth/Kimi-K2.6-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/Kimi-K2.6-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/Kimi-K2.6-GGUF:UD-Q4_K_XL"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use unsloth/Kimi-K2.6-GGUF with Docker Model Runner:
```
docker model run hf.co/unsloth/Kimi-K2.6-GGUF:UD-Q4_K_XL
```

Lemonade

How to use unsloth/Kimi-K2.6-GGUF with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull unsloth/Kimi-K2.6-GGUF:UD-Q4_K_XL

Run and chat with the model

lemonade run user.Kimi-K2.6-GGUF-UD-Q4_K_XL

List all available models

lemonade list

Hermes Agent

How to use unsloth/Kimi-K2.6-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/Kimi-K2.6-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/Kimi-K2.6-GGUF:UD-Q4_K_XL

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use unsloth/Kimi-K2.6-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/Kimi-K2.6-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/Kimi-K2.6-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"

Q4_0 vs native INT4 QAT fidelity

by SpacetimeAI - opened Apr 21

Discussion

SpacetimeAI

Apr 21

Note that llama.cpp’s Q4_0 quantization does not align with the native INT4 QAT format used in post tuning, so converting those tensors to Q4_0 is really just a second quantization step.

In past measurements with k2.5, I saw cosine similarity around ~0.95 between a Q4_0 tensor and the corresponding BF16 tensor, while the Q8 version of that same tensor around ~0.998.

If the quantization algorithm were identical to the original QAT format, I would expect cosine similarity to be near 1.0.

Until llama.cpp implements a quantization scheme that faithfully matches the native INT4 QAT, could we get a Q8 variant that does not use Q4_0 for those tensors

ubergarm

Apr 21

•

edited Apr 21

@SpacetimeAI

I asked about it already here: https://huggingface.co/unsloth/Kimi-K2.6-GGUF/discussions/2#69e796f9db5b3bac09676143

fwiw both myself and AesSedai are using symmetric patch as explained here: https://huggingface.co/ubergarm/Kimi-K2.6-GGUF#q4_x-patch

tl;dr; Q4_X is as close to original int4 as we can get in GGUF land.

ubergarm

Apr 21

I asked over here too: https://www.reddit.com/r/LocalLLaMA/comments/1srxfdh/comment/ohichap/

danielhanchen

Unsloth AI org Apr 22

•

edited Apr 22

Hey sorry for the delay - we wrote about it in https://unsloth.ai/docs/models/kimi-k2.6

Also copying from Ubergarm's other comment below:

Yes we utilized the bijection patch from https://github.com/jukofyork - if not the stock Q4_0 has around 1.8% amax error - now it's within machine epsilon.

Note though Kimi uses BF16 for other tensors, and UD-Q8_K_XL does BF16 for others - we found there does exist some error in the Q4_X variants (UD-Q4_K_XL as well follows Q8_0 for other layers) vs the truly "lossless" UD-Q8_K_XL one seen below:

CC: @SpacetimeAI

SpacetimeAI

Apr 22

Phenomenal work by the community and Unsloth!

ubergarm

Apr 22

Yeah the error isn't bad for Q8_0 especially since the "only 10GB bigger" for bf16 is in the always active tensors for an A32B model users will feel that slow down in TG for sure. Q4_X for the win!

Oh man more Qwen3.6 today already haha, catch you on the next one! Cheers!

shimmyshimmer

Unsloth AI org Apr 22

Yeah the error isn't bad for Q8_0 especially since the "only 10GB bigger" for bf16 is in the always active tensors for an A32B model users will feel that slow down in TG for sure. Q4_X for the win!

Oh man more Qwen3.6 today already haha, catch you on the next one! Cheers!

Yes I agree but Q4_X isn't as close to the original int4 as you you said as it's not lossless unlike the Q8 one.

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