Instructions to use maldv/Qwentile2.5-32B-Instruct with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use maldv/Qwentile2.5-32B-Instruct with Transformers:

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

pipe = pipeline("text-generation", model="maldv/Qwentile2.5-32B-Instruct")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("maldv/Qwentile2.5-32B-Instruct")
model = AutoModelForCausalLM.from_pretrained("maldv/Qwentile2.5-32B-Instruct")
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]:]))

Inference
Notebooks
Google Colab
Kaggle
Local Apps

vLLM

How to use maldv/Qwentile2.5-32B-Instruct with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "maldv/Qwentile2.5-32B-Instruct"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "maldv/Qwentile2.5-32B-Instruct",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/maldv/Qwentile2.5-32B-Instruct

SGLang

How to use maldv/Qwentile2.5-32B-Instruct 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 "maldv/Qwentile2.5-32B-Instruct" \
    --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": "maldv/Qwentile2.5-32B-Instruct",
		"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 "maldv/Qwentile2.5-32B-Instruct" \
        --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": "maldv/Qwentile2.5-32B-Instruct",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use maldv/Qwentile2.5-32B-Instruct with Docker Model Runner:
```
docker model run hf.co/maldv/Qwentile2.5-32B-Instruct
```
Browse Quantizations to use this model in llama.cpp, Ollama, LM Studio, or any compatible app.

mradermacher GGUF quants mradermacher GGUF imat quants Bartowski GGUF quants waldie exl2 4bpw

Qwentile 2.5 32B Instruct

Qwentile 2.5 32B Instruct is a normalized denoised fourier interpolation of the following models:

output_base_model: "Qwen/Qwen2.5-32B"
finetune_merge:
  - { "model": "AiCloser/Qwen2.5-32B-AGI", "base": "Qwen/Qwen2.5-32B", "alpha": 0.3 }
  - { "model": "EVA-UNIT-01/EVA-Qwen2.5-32B-v0.2", "base": "Qwen/Qwen2.5-32B", "alpha": 0.7 }
  - { "model": "fblgit/TheBeagle-v2beta-32B-MGS", "base": "Qwen/Qwen2.5-32B", "alpha": 0.6 }
  - { "model": "huihui-ai/Qwen2.5-32B-Instruct-abliterated", "base": "Qwen/Qwen2.5-32B-Instruct", "alpha": 1.0 }
  - { "model": "huihui-ai/QwQ-32B-Preview-abliterated", "base": "Qwen/Qwen2.5-32B", "alpha": 1.0 }
  - { "model": "Qwen/QwQ-32B-Preview", "base": "Qwen/Qwen2.5-32B", "alpha": 0.8, "is_input": true }
  - { "model": "rombodawg/Rombos-LLM-V2.5-Qwen-32b", "base": "Qwen/Qwen2.5-32B", "alpha": 1.0, "is_output": true }
  - { "model": "nbeerbower/Qwen2.5-Gutenberg-Doppel-32B", "base": "Qwen/Qwen2.5-32B-Instruct", "alpha": 0.4 }

In other words, all of these models get warped and interpolated in signal space, and then jammed back on top of the base model.

What is this?

I started my experiment because of QwQ is a really nifty model, but it was giving me problems with xml output - which is what I use for my thought tokens. So, I thought... lets just merge it in!

The first model worked pretty well, but I got a sense that the balances could be tweaked. Why not throw in some other models as well for fun and see if I can't run out of disk space in the process?

Initial Results

It's a little crispier than Awqward, but does generate stable output. Since it is based on Qwen2.5 base instead of instruct it did not fail the math test, it scores with models twice it's size:

How to think

This model is very compliant to steering, and has innate chain of thought, so producing nice, formatted, chain of thought results is quite easy. Below is a very simple proof-of-concept example of how to achieve a thought turn.

System Prompt

You are my assistant. You should be helpful and use <thought></thought> xml.

All responses are in xml format.

One Shot

<thoughts>
  <thought>Example thought</thought>
  <idea>I have an idea</idea>
</thoughts>

Request

Think through how you render 2d polys at different scales.

Response

<response>
  <thoughts>
    <thought>When rendering 2D polygons at different scales, it's important to maintain their proportions and visual quality.</thought>
    <idea>One approach could be using texture atlases for efficient scaling without losing detail.</idea>
    <thought>Another idea is implementing level-of-detail (LOD) techniques to adjust polygon complexity based on scale.</thought>
    <idea>We might also consider using shaders to handle dynamic scaling and apply effects smoothly.</idea>
  </thoughts>
</response>

I did notice it sometimes drops trailing tags, so you should always validate (and if you are clever, repair) any structured responses.

Citation

If you find our work helpful, feel free to give us a cite.

@misc{qwentile2.5-32b-instruct,
    title = {Qwentile 2.5 32B Instruct},
    url = {https://huggingface.co/maldv/Qwentile2.5-32B-Instruct},
    author = {Praxis Maldevide},
    month = {December},
    year = {2024}
}