Qwen3.6-35B-A3B Claude 4.7 Opus Reasoning Distilled

Qwen3.6-35B-A3B Claude 4.7 Opus Reasoning Distilled

Thinks in explicit think blocks before it answers, tuned on about 8k reasoning traces from Claude Opus 4.7. A 35B mixture-of-experts Qwen3.6 variant for hard math, science, code and planning.

Made by lordx64 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: A community fine-tune of Qwen3.6-35B-A3B that learned to reason in the style of Claude Opus 4.7. The author trained it on about 7,800 reasoning conversations produced by that model, so it writes out its thinking inside think tags before the final answer. It is not an Anthropic release.

Who it suits: People who run their own models for hard reasoning, such as graduate-level science, competition math, code walk-throughs, logic puzzles and agent planning. It fits those who want open weights rather than a proprietary API.

Industry | Who uses it | How they use it | Example
Education | Math and science students | Work through hard problems with visible reasoning | A step-by-step solution to a competition problem
Software | Developers | Ask for code reasoning with an explicit walk-through | Tracing why a function fails
Research | AI researchers | Study reasoning distillation into an open model | Comparing it with its base model
Agent building | Agent developers | Plan multi-step tasks where thinking helps accuracy | A planner that reasons before choosing a tool

You can skip it if you need fast, short chat replies, since the thinking can run to many thousands of tokens.

Where to use it:

• Python with the transformers library
• vLLM, which the author recommends for serving
• LM Studio and llama.cpp, through the separate GGUF copies
• A GPU with 80 GB, such as an A100 or H100, for full bf16

Use case | Who | What they get back
Hard reasoning questions | Students and researchers | A long think block, then a final answer
Serving an open reasoning model | Developers | An endpoint that streams reasoning and answers
Local chat | Enthusiasts | A quantized copy in LM Studio

How to set it up:

• Python on Windows, macOS or Linux
• The huggingface_hub package, and transformers or vLLM to run it
• About 71.9 GB for the BF16 weights, or about 18 GB at 4-bit by Hugging Face's estimate

1. Install the Hugging Face client.
2. Download the model files to your computer.
3. Load the model with transformers, or serve it with vLLM.

pip install -U huggingface_hub
hf download lordx64/Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled

How to use it: Send a chat message and let the model finish its thinking before it answers. Lower max_new_tokens to shorten the output. If your app shows only the answer, remove the think blocks afterwards.

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

repo = "lordx64/Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(
    repo, torch_dtype=torch.bfloat16, device_map="auto", trust_remote_code=True,
)

messages = [{"role": "user", "content": "How many positive integers less than 1000 have digits that sum to 20?"}]
inputs = tok.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)
out = model.generate(inputs, max_new_tokens=32768, do_sample=False)
print(tok.decode(out[0][inputs.shape[-1]:], skip_special_tokens=True))

To serve it, the card recommends vLLM.

vllm serve lordx64/Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled \
  --dtype bfloat16 --max-model-len 65536 --gpu-memory-utilization 0.9

Tool | What it does
transformers | Loads the model and runs generation
vllm serve | Serves the model with continuous batching
max_new_tokens | Caps how long the thinking and answer can run
GGUF copies | Quantized files for LM Studio and llama.cpp

Example: You ask how many three-digit numbers have digits that add up to 15. The model writes a long think block that works through the cases, then gives a short final answer. You can check the steps it showed before trusting the number.

For newcomers and experts: If you are new, try a quantized GGUF copy in LM Studio first, since it needs far less memory. Ask a short math question and watch the thinking appear before the answer.

If you know the field: the base has 256 experts with 8 routed and 1 shared, and about 3B parameters active per token. The LoRA fine-tune targets only the attention projections and trains the loss on assistant tokens, think blocks included.

Claude Opus 4.7 reasoning traces
SFT on Qwen3.6-35B-A3B
Think block then answer
64k context

Good to know: The author lists three GGUF sizes in separate repositories, from a small default pick to a near-lossless one. A LoRA adapter is also published on its own.

Title: GGUF file sizes in GB
IQ4_XS: 18.9 GB
Q5_K_M: 25 GB
Q8_0: 35 GB

• Context: 64k tokens, with thinking that often runs 5k to 30k tokens on hard problems
• Training data: about 7,800 conversations from Claude Opus 4.7, reformatted for fine-tuning
• License: Apache 2.0
• Cost to run: none named
• Data it reaches: model files downloaded to your computer, nothing else named
• Not for: latency-sensitive short chat, since the thinking budget can be large

On Hugging Face, 13 November 2026

• Task: text generation
• Library: transformers
• Parameters: 36B
• Weights: about 71.9 GB in BF16; about 18 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 674
• Likes: 206
• License: apache-2.0
• Languages: en
• Based on: Qwen/Qwen3.6-35B-A3B
• Quantized versions: 47
• Fine-tunes: 5
• Tags: transformers, safetensors, qwen3_5_moe, image-text-to-text, text-generation, reasoning, distillation, chain-of-thought, qwen, qwen3.6, mixture-of-experts, lora

Links

• Model: https://huggingface.co/lordx64/Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled
• Files: https://huggingface.co/lordx64/Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download lordx64/Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled

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