Qwen3.8-9B-Distill

Qwen3.8-9B-Distill

Reasons step by step through math, code and general questions, opening each answer with a thinking block. A 9B Qwen3.5 model trained on traces from a larger Qwen3.8 teacher, for a single GPU.

Made by empero-ai on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Qwen3.8-9B-Distill is a 9B model from Empero. It is a full parameter distillation of a much larger Qwen3.8 teacher into the Qwen3.5-9B architecture. The student was trained on about 70,000 filtered teacher answers that include their reasoning. The mix leans toward hard maths and competitive programming, with some everyday and tool use cases. The goal is to bring a large model's reasoning into a dense model that fits on one GPU.

Who it suits: Developers, students and researchers who want a reasoning model they can run on a single GPU.

Industry | Who uses it | How they use it | Example
Software development | Developers who run local models | Ask coding and tool use questions with reasoning shown | Asking for a fix to a failing function call
Education | Students of maths and logic | Work through word problems with the thinking visible | Solving a puzzle about a snail climbing a well
Research | People who study distillation | Compare the student with its Qwen3.5-9B base | Running the same benchmark on both models

Skip it if you need short instant answers, since every reply opens with a thinking block.

Where to use it: Any runtime that supports the Qwen3.5 architecture should load it, including Hugging Face transformers, vLLM and SGLang. The files download on Windows, macOS or Linux.

• transformers, vLLM or SGLang
• A GPU with room for about 19.3 GB of BF16 weights, or about 4.8 GB at 4-bit as an estimate
• Windows, macOS or Linux: where you can fetch the files with Python

Use case | Who | What they get back
Math and logic | Students | A thinking block and a final answer
Code questions | Developers | Reasoned explanations and code
Function calling | Agent builders | Tool calls in Qwen3.5's native format

How to set it up: Start with Python and a GPU. Your transformers release must be recent enough to include Qwen3.5.

• Python with pip
• A recent transformers release
• Two extra kernel packages, flash-linear-attention and causal_conv1d, without which some layers run slowly and use more memory

1. Install the Hugging Face hub library.
2. Download the model files.
3. Install a recent transformers and the Gated DeltaNet kernels.
4. Run the quickstart code.

pip install -U huggingface_hub
hf download empero-ai/Qwen3.8-9B-Distill

How to use it: Load the model, send a chat message and allow plenty of new tokens. The card recommends a temperature of 0.6, top_p of 0.95 and top_k of 20, and warns that greedy decoding on long answers can loop.

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "empero-ai/Qwen3.8-9B"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, dtype=torch.bfloat16, device_map="auto")

messages = [{"role": "user", "content": "A snail is at the bottom of a 10-meter well. Each day it climbs 3 meters, each night it slips back 2. How many days until it escapes?"}]
inputs = tok.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)

out = model.generate(inputs, max_new_tokens=16384,
                     temperature=0.6, top_p=0.95, top_k=20, do_sample=True)
print(tok.decode(out[0][inputs.shape[1]:], skip_special_tokens=True))

Tool | What it does
apply_chat_template | Builds the chat prompt for the model
model.generate | Produces the answer, with up to 16,384 new tokens here

Example: You ask the snail and well riddle from the quickstart. The model writes a thinking block that works through the climb and slip pattern, then gives the number of days.

For newcomers and experts: If you are new to this, expect a long thinking section before the answer. Strip the think block when you show the reply to other people.

If you know the field, the student was trained by supervised fine tuning on teacher traces and not on its own attempts. So it inherits the teacher's style, including occasional long deliberation on easy questions. The fine tune is text only, and vision behaviour from the base was not tested.

Large Qwen3.8 teacher writes traces
About 70,000 traces filtered
Qwen3.5-9B student fine tuned
You run it on one GPU

Good to know: The card reports its own tests against the Qwen3.5-9B base, using chain of thought prompts. Scores rose on MMLU and fell slightly on GSM8K.

Title: Card test, flexible match scores
Task | Qwen3.5-9B base | Qwen3.8-9B
MMLU with chain of thought | 0.546 | 0.751
GSM8K with chain of thought | 0.885 | 0.870

• License: Apache 2.0
• Cost to run: none named, you supply the hardware
• Data it reaches: the card names none
• Not for: image tasks, since the fine tune is text only

On Hugging Face, 14 November 2026

• Task: text generation
• Library: transformers
• Parameters: 9.7B
• Weights: about 19.3 GB in BF16; about 4.8 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 6,592
• Likes: 237
• License: apache-2.0
• Languages: en
• Based on: Qwen/Qwen3.5-9B
• Quantized versions: 38
• Fine-tunes: 13
• Tags: transformers, safetensors, qwen3_5, image-text-to-text, empero-ai, qwen3.5, qwen3.8, distillation, reasoning, function-calling, text-generation, conversational

Links

• Model: https://huggingface.co/empero-ai/Qwen3.8-9B-Distill
• Files: https://huggingface.co/empero-ai/Qwen3.8-9B-Distill/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download empero-ai/Qwen3.8-9B-Distill

Then follow the usage in its model card. Nothing in this kit runs by itself.