Hello everyone. Finally I fixed Qwen3.5 35B A3B model after finetuning on Claude Opus datasets and merging with HauhauCS uncensored version. I fixed: attention, Dense FFN, MoE routed experts, and MoE shared expert tensors in model.

Here link: https://huggingface.co/LuffyTheFox/Qwen3.5-35B-A3B-Claude-4.6-Opus-Uncensored-KL-UD-V2-GGUF . Please read launch instructions on page for best experience.

I merged: samuelcardillo model with HauhauCS model, and applied my fixes.
Merging has been done via this script: https://pastebin.com/eB6zB4DU

Model programming features has been tested via following prompts:

1. Write an Arkanoid game using HTML5 and Javascript. The game should be controlled with a mouse and include generated sounds and effects. The game should have beautiful design with neon bricks and sounds.
2. Add bonus system. Change background to space.

I got this result: https://pastebin.com/P29JEnPA

Bonus script: Universal Dynamic quantization workflow for Google Colab Free (CPU).

Quantization has been done via this script for UD Q4_K_XL quant: https://pastebin.com/5Ba6qs7L

My idea:

1. Read the exact per-tensor quantization types used in: Qwen3.5-35B-A3B-UD-Q4_K_XL.gguf (Unsloth) quant.
2. Save them into a unsloth_ud_profile.json here link: https://pastebin.com/qYrFYadc
3. Delete Unsloth reference quant to save disk.
4. Quantize your finetuned GGUF (Q8_0/BF16) -> Q4_K_XL using that JSON profile.

Enjoy \^_\^. Please upvote / share if you like the model and my scripts.