Qwen3.8-35B-A3B-Distill

Qwen3.8-35B-A3B-Distill

Reasons step by step before it answers, using a 35B mixture-of-experts model that activates about 3B parameters per token. Distilled from Qwen3.8 teacher traces, with a focus on math and code.

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: Qwen3.8-35B-A3B-Distill is a smaller student model from Empero, trained on worked answers from the larger Qwen3.8 models. It sits on the Qwen3.6-35B-A3B architecture, which holds 35B parameters but uses about 3B per token, with 8 of 256 experts routed each time. Every reply opens with a think block that shows the reasoning.

Who it suits: It suits developers and researchers who want a reasoning model that aims to run on one GPU and is strong on math and code.

Industry | Who uses it | How they use it | Example
Software development | Developers | Run a reasoning model locally for code questions | Solving a programming puzzle step by step
Education | Tutors and students | Work through math word problems with the reasoning shown | Explaining a puzzle about a snail in a well
Agent building | Agent developers | Use its built-in function calling in tool workflows | Letting the model call a tool from a chat
Research | ML researchers | Study distillation from teacher traces | Comparing it with its base model on ARC

You can skip it if you need a small download, since the weights are about 70 GB in BF16.

Where to use it:

• Hugging Face Transformers in Python
• vLLM, SGLang and other runtimes that support the Qwen3.6 architecture
• Windows, macOS or Linux with Python, to download the files

Use case | Who | What they get back
Math and code questions | Students and developers | A reasoning block followed by an answer
Tool use | Agent developers | Function calls that follow the Qwen3.6 specification
Long inputs | Analysts | Handling of context up to 262,144 tokens

How to set it up: Plan for Python, a GPU that can hold the weights, and an up-to-date transformers install that knows the Qwen3.6 architecture.

• Python and huggingface_hub for the download
• A recent transformers version that supports Qwen3.6
• Two kernel add-ons: flash-linear-attention and a causal_conv1d build matched to your CUDA version, since plain PyTorch fallbacks run slower and use more memory
• Roughly 70 GB for BF16 weights, or about 17.6 GB at 4-bit as an estimate

1. Install huggingface_hub.
2. Download the model files.
3. Install a recent transformers and the two kernel packages.
4. Run the quickstart script.

pip install -U huggingface_hub
hf download empero-ai/Qwen3.8-35B-A3B-Distill

How to use it: The card's quickstart loads the model, sends one chat message and prints the reply. Its model id is spelled Distilled, so use the name of the repository you downloaded.

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "empero-ai/Qwen3.8-35B-A3B-Distilled"
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", return_dict=True).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["input_ids"].shape[1]:], skip_special_tokens=True))

Tool | What it does
temperature, top_p, top_k | Sampling values suggested by the authors: 0.6, 0.95 and 20
max_new_tokens 16384 | Leaves room for the think block and the answer
do_sample=True | Avoids greedy decoding, which can loop on long outputs
AutoModelForCausalLM | Loads the text path of the model

Example: You send the snail-in-a-well riddle from the quickstart. The model writes a think block with its reasoning, then the final answer. You strip the think block before showing the reply to end users.

For newcomers and experts: If you are new to this, distillation means a smaller model learns from the worked answers of bigger ones. A first small try is the quickstart script, then reading the reasoning before the answer.

For people who know the field, training was supervised and off-policy on curated teacher traces, and it updated the attention path as well as the routed and shared expert stacks. The base uses 40 layers with hybrid linear and full attention. The card reports ARC gains outside noise and an MMLU change within noise.

Title: ARC acc_norm, base model against this model
ARC-Challenge base: 0.548
ARC-Challenge this model: 0.591
ARC-Easy base: 0.717
ARC-Easy this model: 0.766

Teacher traces from Qwen3.8 models
Supervised distillation into the 35B MoE student
Your prompt goes through the chat template
The model writes a think block, then the answer

Good to know: The benchmarks use zero-shot loglikelihood scoring with the same settings for base and student, so they say little about long reasoning chains.

• License: Apache 2.0
• Cost to run: none named
• Data it reaches: the prompts you send and the downloaded model files
• Not for: greedy decoding on long generations, which the card says can fall into repetition loops

On Hugging Face, 13 November 2026

• Task: text generation
• Library: transformers
• Parameters: 35.1B
• Weights: about 70.2 GB in BF16; about 17.6 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 4,390
• Likes: 103
• License: apache-2.0
• Languages: en
• Based on: Qwen/Qwen3.6-35B-A3B
• Quantized versions: 23
• Fine-tunes: 1
• Tags: transformers, safetensors, qwen3_5_moe, image-text-to-text, empero-ai, qwen3.6, qwen3.8, distillation, reasoning, function-calling, text-generation, conversational

Links

• Model: https://huggingface.co/empero-ai/Qwen3.8-35B-A3B-Distill
• Files: https://huggingface.co/empero-ai/Qwen3.8-35B-A3B-Distill/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download empero-ai/Qwen3.8-35B-A3B-Distill

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