Instructions to use unsloth/Kimi-K3 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use unsloth/Kimi-K3 with Transformers:

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

pipe = pipeline("image-text-to-text", model="unsloth/Kimi-K3", trust_remote_code=True)
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-K3", trust_remote_code=True, device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use unsloth/Kimi-K3 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "unsloth/Kimi-K3"
# 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-K3",
		"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-K3

SGLang

How to use unsloth/Kimi-K3 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-K3" \
    --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-K3",
		"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-K3" \
        --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-K3",
		"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"
						}
					}
				]
			}
		]
	}'

Unsloth Studio

How to use unsloth/Kimi-K3 with Unsloth Studio:

Install Unsloth Studio (macOS, Linux, WSL)

curl -fsSL https://unsloth.ai/install.sh | sh
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for unsloth/Kimi-K3 to start chatting

Install Unsloth Studio (Windows)

irm https://unsloth.ai/install.ps1 | iex
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for unsloth/Kimi-K3 to start chatting

Using HuggingFace Spaces for Unsloth

# No setup required
# Open https://huggingface.co/spaces/unsloth/studio in your browser
# Search for unsloth/Kimi-K3 to start chatting

Load model with FastModel

pip install unsloth
from unsloth import FastModel
model, tokenizer = FastModel.from_pretrained(
    model_name="unsloth/Kimi-K3",
    max_seq_length=2048,
)

Docker Model Runner
How to use unsloth/Kimi-K3 with Docker Model Runner:
```
docker model run hf.co/unsloth/Kimi-K3
```

System Requirements

by TheEnlightenedMalignancy - opened 3 days ago

Discussion

TheEnlightenedMalignancy

3 days ago

Hi, can I run this on my 2080ti?

xsrrr

3 days ago

yes and no. You'd need the 7-12B parameter version to be able to run it on 2080ti, but then, it won't be k3, right? If you dumb it enough....

rallymoose

3 days ago

How much vram do you have? and how much ram, that's the two most deciding factors.
look at:
https://github.com/chrishayuk/larql
https://github.com/JustVugg/colibri

None of them can run kimi k3. But that's a tip to at least run bigger models.

But one last tip... Get a newer GPU.

CYISNOTHERE

3 days ago

•

edited 3 days ago

The moment I saw that model size, all of a sudden I started hearing Attack on Titan music...👀

attack ON titan - Hiroyuki Sawano

invincibilul

2 days ago

•

edited 2 days ago

you need at least 2000 gb of ram to run this lmao , very few people will be able to run this, but its worth having since USA will ban open source AI

CYISNOTHERE

2 days ago

you need at least 2000 gb of ram to run this lmao , very few people will be able to run this, but its worth having since USA will ban open source AI

Time to use VPNs like us euro poors 😂

Ayylmayonnaise

2 days ago

•

edited 2 days ago

You technically can.
Some one managed to run it on M1 Mac 96GB at 3 tokens per minute :D
https://old.reddit.com/r/LocalLLM/comments/1v8jfo2/got_kimi_k3_running_on_my_macbook_its_painfully/

Would be something like 1 token per 10 minutes or something in your case.

invincibilul

2 days ago

vpns wont help, they will just stop huggingface, you will have to find it on torrents if USA goes that way, regardless, its best to save it on an ssd.

CYISNOTHERE

2 days ago

vpns wont help, they will just stop huggingface, you will have to find it on torrents if USA goes that way, regardless, its best to save it on an ssd.

huggingface is french with its HQ in the US, at best it might create a need to split huggingface between US and the rest of the world. At worse if they can't move to non-US servers then yeah the scenario you said will happen. Might spur the creation of more huggingface type startups globally.

invincibilul

2 days ago

i love your childish view of the internet and political , but the real world doesn't work like that

SirDigby

2 days ago

Here's about how much it'll cost.

https://anafrashop.com/nvidia-baseboard-8x-b300-935-26287-0070-000-5

CYISNOTHERE

2 days ago

i love your childish view of the internet and political , but the real world doesn't work like that

jortony

1 day ago

I priced a capable single node system from SuperMicro at about $3M USD: 5.8TB of DDR5 and 8xB300

CYISNOTHERE

1 day ago

I priced a capable single node system from SuperMicro at about $3M USD: 5.8TB of DDR5 and 8xB300

anyone got $3M lying around that I can borrow? 👀

invincibilul

1 day ago

i love your childish view of the internet and political , but the real world doesn't work like that

your reply just proved my point.

CYISNOTHERE

1 day ago

i love your childish view of the internet and political , but the real world doesn't work like that

your reply just proved my point.

Naive optimism is the only way to keep moving forward in this depressing world, you can try to be a bit more hopeful over the future of chinese opensource otherwise you wouldn't be here replying 🤷‍♂️

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