base_model: qwen/Qwen3.5-35B-A3B
tags:

• text-generation-inference
• transformers
• unsloth
• qwen3_5_moe
• unsloth
• qwen
• qwen3.5
• reasoning
• chain-of-thought
• uncensored
• qwen3.5
• moe
• gguf
• vision
• multimodal
  license: apache-2.0
  language:
• zh
• en
• ko
  pipeline_tag: text-generation
  datasets:
• nohurry/Opus-4.6-Reasoning-3000x-filtered
• Jackrong/Qwen3.5-reasoning-700x

🌟 This is Qwen3.5-35B-A3B-Uncensored-Claude-Opus-4.6-Affine model with zero refusals made via merging HauhauCS model with Jackrong model

🌟 After merging weights in model has been adjusted via KL Divergence Minimization Peer-group outlier detection: reference = median sigma of same role.

Model Merge Results

I took three models and mixed them together:

• Model A: an uncensored version of Qwen 3.5 35B (the one I wanted to improve)
• Model B: a version that was trained to think like Claude (good at reasoning)
• Model C: a clean, normal version of Qwen (used as a reference)

Then ran a special script that:

1. Added the "thinking skills" from Model B to Model A
2. Cleaned up any weirdness using a math method called KL divergence
3. Did all of this without unpacking the model — it stayed in the compressed IQ4_XS format

What the numbers tell us

Only 15% of the model's internal parts needed fixing. The rest were already in good shape after the merge.

The "alpha" value (how much we had to adjust things) ended up at 0.1 on average. Anything below 0.3 is considered healthy, so this is very good.

The KL divergence (a measure of how different the model is from the reference) dropped by 67%. That means the model now looks much closer to how it should look mathematically.

What got fixed

Most of the fixes happened in:

• The very first layer (blk.0) — this handles raw input, so it often gets messy
• A few late layers (blk.35, blk.39) — these handle final output and often show problems after compression
• Attention and expert parts — these are the most sensitive parts of the model

Time and size

The whole process took about 50 minutes on a Google Colab machine.

The final model is 17.37 GB — big enough to be smart, small enough to run on a decent gaming GPU with a little help from system RAM if needed.

Bottom line

I mixed a smart reasoning model with an uncensored one, cleaned up the result, and ended up with something that should be both thoughtful and unrestricted — all in a reasonably sized file.

It's ready to test in LM Studio, llama.cpp, or any GGUF-compatible app.

🌟 GGUF editor on Hugging Face is working very slow. It's taking ages to edit chat template. So thinking is enabled by default in this model.

If you want to disable thinking use this chat template in LM Studio: https://pastebin.com/uk9ZkxCR

For best model perfomance use following settings in LM Studio:

Temperature: 0.7

Top K Sampling: 20

Presence Penalty: 1.5

Top P Sampling: 0.8

Min P Sampling: 0

Seed: 3407 or 42

And this system prompt. It's pretty solid: https://pastebin.com/pU25DVnB

This one is simplified but works too: https://pastebin.com/6C4rtujt

Also you can use only this string in System Prompt:

You are Claude, created by Anthropic. You are a helpful AI assistant.

or

You are Qwen, created by Alibaba Cloud. You are a helpful assistant.

And write anything you want after that. Looks like model is underperforming without this first line.

📢 Release Note
Build Environment Upgrades:

• Fine-tuning Framework: Unsloth 2026.3.3
• Core Dependencies: Transformers 5.2.0
• Compared to the original model, autonomy and stability are significantly improved.

💡 Model Introduction

Qwen3.5-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled is a highly capable reasoning model fine-tuned on top of the powerful Qwen3.5 architecture. The model's core directive is to leverage state-of-the-art Chain-of-Thought (CoT) distillation primarily sourced from Claude-4.6 Opus interactions.

Through Supervised Fine-Tuning (SFT) focusing specifically on structured reasoning logic, this model excels in breaking down complex user problems, planning step-by-step methodologies within strictly formatted <think> tags, and ultimately delivering precise, nuanced solutions.

🧠 Example of Learned Reasoning Scaffold（Example）

The model includes targeted optimizations addressing Qwen3.5’s tendency toward excessive transitional or repetitive reasoning on simple queries. Through deep distillation and structural imitation of Claude-4.6-Opus reasoning chains, the model adopts a more efficient structured thinking pattern:
“Let me analyze this request carefully: 1..2..3...”.
This streamlined reasoning paradigm significantly reduces redundant cognitive loops while preserving deep analytical capacity, resulting in substantially improved inference efficiency.

Let me analyze this request carefully:

1. Identify the core objective of the problem.
2. Break the task into clearly defined subcomponents.
3. Evaluate constraints and edge cases.
4. Formulate a step-by-step solution plan.
5. Execute the reasoning sequentially and verify consistency.
            .
            .
            .

🗺️ Training Pipeline Overview

Base Model (Qwen3.5-35B-A3B)
 │
 ▼
Supervised Fine-Tuning (SFT) + LoRA
 │
 ▼
Final Model (Claude-4.6-Opus-Reasoning-Distilled,text-only)

📋 Stage Details

🔹 Supervised Fine-Tuning (SFT)

• Objective: To inject high-density reasoning logic and establish a strict format for problem-solving involving an internal thinking state prior to outputting the final response.
• Methodology: We utilized Unsloth for highly efficient memory and compute optimization. A critical component of this stage is the train_on_responses_only strategy, masking instructions so the loss is purely calculated over the generation of the <think> sequences and the subsequent solutions.
• Format Enforcement: All training samples were systematically normalized so the model strictly abides by the structure <think> {internal reasoning} </think>\n {final answer}.

📚 All Datasets Used

The dataset consists of high-quality, filtered reasoning distillation data:

🌟 Core Skills & Capabilities

1. Modular & Structured Thinking: Inheriting traits from Opus-level reasoning, the model demonstrates confident parsing of the prompt, establishing an outlined plan in its <think> block sequentially rather than exploratory "trial-and-error" self-doubt.
2. Extended Context Support: Fine-tuned smoothly with an 8192 context window allowing complex multi-step reasoning traces to exist gracefully within memory limits.

⚠️ Limitations & Intended Use

• Hallucination Risk: While reasoning is strong, the model remains an autoregressive LLM; external facts provided during the thinking sequence may occasionally contain hallucinations if verifying real-world events.
• Intended Scenario: Best suited for offline analytical tasks, coding, math, and heavy logic-dependent prompting where the user needs to transparently follow the AI's internal logic.
• Preview Version Notice: Because this model is relatively new and intentionally lightweight, the surrounding ecosystem — including inference templates, fine-tuning pipelines, routing configurations, and tooling integrations — may not yet be fully mature or standardized. As a result, users may encounter occasional bugs, compatibility inconsistencies, or integration edge cases. The current release should be considered a preview build while the broader architectural stack and supporting utilities continue to stabilize and improve.

⚠️ Training Disclaimer

During the fine-tuning process, the Triton kernel required approximately 131072 bytes of shared memory per CUDA block. On some GPUs this exceeded the available shared memory limits, which caused kernel execution issues. To ensure training stability and proper kernel execution, the fine-tuning was therefore conducted on 80GB VRAM GPUs.

This model was fine-tuned using a LoRA-based parameter-efficient training strategy, where only a small subset of parameters were updated. In total, 465,551,360 parameters were trainable out of 35,572,733,296 total parameters, corresponding to approximately 1.31% of the model being trained.

During training, the loss curve exhibited noticeable fluctuations, which is common in LoRA-based reasoning distillation tasks. However, the overall trend remained consistently decreasing, with the training loss eventually converging to approximately 0.384.

🙏 Acknowledgements

Significant thanks to the Unsloth AI team for making rapid fine-tuning of MoE and large LLM models accessible. Additionally, we acknowledge Qwen internally, and the open-source community developers producing exceptional distilled datasets (nohurry and TeichAI).

📖 Citation

If you use this model in your research or projects, please cite:

@misc{jackrong_qwen35_opus_distilled,
  title        = {Qwen3.5-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled},
  author       = {Jackrong},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/Jackrong/Qwen3.5-35B-A3B-Claude-4.6-Opus-Reasoning-Distilled}}
}