Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled-GGUF

Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled-GGUF

Runs a reasoning-tuned Qwen3.6-35B-A3B on your own machine as GGUF files in four quant sizes, trained on Claude Opus 4.6 style reasoning data. For people who want step-by-step answers locally.

Made by hesamation 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: These are GGUF quantizations of a Qwen3.6-35B-A3B model that was fine-tuned for reasoning. The training data was chain-of-thought material in the style of Claude Opus 4.6. The files cover text only, so image and video input is not part of what was trained.

Who it suits: It suits people who want to run a reasoning-focused model on their own computer and compare it with its base model.

Industry | Who uses it | How they use it | Example
Local AI | Hobbyists with strong hardware | Run a 35B reasoning model offline | Working through a logic puzzle step by step
Research | Fine-tuning researchers | Study the effect of reasoning distillation on a base model | Comparing answers with the original Qwen model
Software | Developers | Try a reasoning model in a local llama.cpp setup | Asking a technical question and reading the working
Education | Students and tutors | Read worked reasoning on problems | Walking through a maths question

Skip it if you need image input or have little memory, since the 4-bit weights alone are about 17.3 GB by estimate.

Where to use it:

• llama.cpp and other runtimes that read GGUF files
• Windows, macOS or Linux with Python, for downloading the files
• Local chat or scripting setups where you pick the quant that fits your memory

Use case | Who | What they get back
Reasoning questions | Local AI users | Answers that show their working
Model comparison | Researchers | A tuned model to test against the base Qwen model
Quant choice | Anyone tuning memory use | Four file sizes to trade quality against memory

How to set it up: You need these first.

• Python with pip
• A runtime that loads GGUF files, such as llama.cpp
• Enough memory for your chosen quant, starting near 17.3 GB for 4-bit

1. Install the Hugging Face hub library with pip.
2. Download the model files with the hf command.
3. Pick one quant file, since the card says files arrive as each quantization finishes.
4. Load that file in your GGUF runtime, following its own instructions.

pip install -U huggingface_hub
hf download hesamation/Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled-GGUF

How to use it: Load one quant file in your runtime and chat with it as a reasoning model. The training used the qwen3-thinking chat template. The card gives no run command of its own, so the exact steps depend on your runtime.

Tool | What it does
Q4_K_M | The smallest everyday choice for local use
Q5_K_M | A little larger, with better quality than Q4
Q6_K | Higher quality when your memory allows
Q8_0 | The largest file, closest to the source model

Example: Load the Q4_K_M file and ask it to solve a word problem and show its steps. You get a written line of reasoning followed by an answer. The card gives no sample output, so expect the style to vary with your settings.

For newcomers and experts: If you are new, a quant is a compressed copy of a large model. Start with Q4_K_M because it is the smallest, and try one reasoning question before changing anything else.

If you know the field, this is a LoRA fine-tune on attention modules only, rank 32 and alpha 32, run for 2 epochs and 762 steps, then merged and quantized with llama.cpp. Three reasoning datasets fed the training. The one benchmark is a small MMLU-Pro smoke test on the merged source model, not on each GGUF, and the author calls it a comparative check, not a full benchmark.

Qwen3.6-35B-A3B base model
Supervised fine-tune with LoRA
Merge into one full model
Quantize to GGUF with llama.cpp
Four quant files to choose from

Good to know:

• Text only: the base design has a vision encoder, but this run did not train on images or video
• Benchmarks: the score was measured on the merged source model, and quantizing can change scores
• Other builds: other people have made 2 fine-tuned versions of it

Title: MMLU-Pro exact match on 70 questions
Base model: 42.86%
Merged source model: 75.71%

• License: Apache 2.0
• Cost to run: none named, the source lists no API key, account or paid service
• Data it reaches: the files you download and the prompts you give your own runtime, and the source names no web or account access
• Not for: image or video input, and not backed by a full-scale benchmark, since the one test is a 70-question check

On Hugging Face, 13 November 2026

• Task: text generation
• Library: gguf
• Parameters: 34.7B
• Weights: about 17.3 GB at 4-bit (estimate)
• Formats: GGUF
• Downloads in the last 30 days: 17,179
• Likes: 296
• License: apache-2.0
• Languages: en
• Based on: hesamation/Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled
• Fine-tunes: 2
• Tags: gguf, llama.cpp, qwen, qwen3.6, qwen3_5_moe, reasoning, chain-of-thought, conversational, quantized, unsloth, text-generation, model-index

Links

• Model: https://huggingface.co/hesamation/Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled-GGUF
• Files: https://huggingface.co/hesamation/Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled-GGUF/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download hesamation/Qwen3.6-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled-GGUF

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