Instructions to use zerofata/MS3.2-PaintedFantasy-Visage-33B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use zerofata/MS3.2-PaintedFantasy-Visage-33B with Transformers:

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

pipe = pipeline("text-generation", model="zerofata/MS3.2-PaintedFantasy-Visage-33B")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("zerofata/MS3.2-PaintedFantasy-Visage-33B")
model = AutoModelForCausalLM.from_pretrained("zerofata/MS3.2-PaintedFantasy-Visage-33B")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use zerofata/MS3.2-PaintedFantasy-Visage-33B with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "zerofata/MS3.2-PaintedFantasy-Visage-33B"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "zerofata/MS3.2-PaintedFantasy-Visage-33B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/zerofata/MS3.2-PaintedFantasy-Visage-33B

SGLang

How to use zerofata/MS3.2-PaintedFantasy-Visage-33B 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 "zerofata/MS3.2-PaintedFantasy-Visage-33B" \
    --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": "zerofata/MS3.2-PaintedFantasy-Visage-33B",
		"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 "zerofata/MS3.2-PaintedFantasy-Visage-33B" \
        --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": "zerofata/MS3.2-PaintedFantasy-Visage-33B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use zerofata/MS3.2-PaintedFantasy-Visage-33B with Docker Model Runner:
```
docker model run hf.co/zerofata/MS3.2-PaintedFantasy-Visage-33B
```

GPT-Like

by poipoi300 - opened Jul 16, 2025

Discussion

poipoi300

Jul 16, 2025

•

edited Jul 19, 2025

It's obvious that this model inherits a lot of qualities from models such as ChatGPT (looking at you, em dash, it's not this but that, etc etc.). If you do not like ChatGPT's writing, then I doubt you'll be a fan of this one. That said, and like others have mentioned, it has potential. It absolutely needs a strong repetition penalty (1.2), DRY (1.6, 2, 5) and a well-written character card. I also have dynamic temp (0.7-2). These are just preliminary settings, but without those thereabouts, it'll carry over some offhand character trait in every message for probably longer than its context length, and they pile up to the point of unending spirals sometimes. Perhaps someone else can research proper sampling for real, but it's not too bad on those mentioned. XTC on top made it weird. It still repeats structure a lot.

Maybe, just maybe, this model would be better than Violet Twilight (and further fine-tunes) in the 12B realm with the right sampling (looking rough for this model being compared with 12B ones, but it's kind of not even close), but I fear it has unsolvable biases to truly provide an enjoyable experience compared even with much smaller models. It has a very uneven intelligence distribution as well. It produces a lot of sentences that have you going "Uh, yeah, makes s- Eh?" regardless of parameters.

Now, I understand this release is experimental, and it seems like others do like it. I saw this model recommended on reddit with a fair amount of upvotes. Ultimately, I can't say I feel the same about this. Now, to go play with my wife a bit more before her lobotomy.

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