license: apache-2.0
library_name: transformers
tags:
- qwen
- qwen3.6
- moe
- multimodal
- vision
- uncensored
- fp16
- safetensors
- transformers
language: - en
base_model: HauhauCS/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive
Qwen3.6-35B-A3B-Uncensored-HauhauCS-FP16
This repository is a HuggingFace reconstruction project and is not an official release from Qwen Team or HauhauCS.
Overview
This repository provides a HuggingFace Transformers compatible FP16 checkpoint reconstructed from the publicly released GGUF model:
HauhauCS/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive
The purpose of this project is to reconstruct a complete HuggingFace checkpoint from the public GGUF release while preserving architecture, tensor structure, metadata, multimodal functionality, tokenizer compatibility, and Multi-Token Prediction (MTP) support as closely as possible.
This repository is intended for:
- Research
- Preservation
- Interoperability
- Compatibility with the HuggingFace ecosystem
- Educational study of model reconstruction techniques
- Long-term archival of publicly released model weights
This repository is not a finetune.
This repository is not a retrained model.
No additional training, RLHF, DPO, SFT, alignment, model merging, pruning, distillation, quantization, censorship, uncensorship, or behavioral modification has been performed.
Source Models
Original Base Model
Qwen Team
Model:
Qwen/Qwen3.6-35B-A3B
Public GGUF Release
HauhauCS
Model:
Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive
All credit for the uncensored release belongs to HauhauCS.
Reconstruction Author
Tiến Khôi Lê
Work performed in this repository includes:
- GGUF tensor analysis
- Tensor reconstruction
- Metadata reconstruction
- HuggingFace checkpoint reconstruction
- Safetensors conversion
- MTP restoration
- Compatibility validation
Reconstruction Methodology
The reconstruction process included:
- GGUF tensor extraction
- Tensor inventory reconstruction
- Quantization block analysis
- Tensor mapping recovery
- Weight reconstruction
- Metadata reconstruction
- HuggingFace configuration reconstruction
- Safetensors shard generation
- Multi-Token Prediction (MTP) restoration
- Transformers compatibility validation
- Checkpoint integrity verification
The goal was to preserve compatibility with the original HuggingFace architecture while maintaining support for advanced model features.
Detailed Dequantization Methodology
A detailed technical description of the dequantization process is available in:
METHOD.md
Topics covered include:
- GGUF quantization formats
- Quantization block structure analysis
- Scale and parameter recovery
- Tensor-level dequantization formulas
- Reconstruction of floating-point weights
- Numerical properties and limitations of dequantized tensors
The document focuses on the mathematical and implementation details required to reconstruct floating-point tensors from the public GGUF release.
Technical Information
Architecture
- Qwen3.6-35B-A3B
- Mixture of Experts (MoE)
- 35B total parameters
- Approximately 3B active parameters per token
- 256 experts
- 8 routed experts per token
- 40 transformer layers
- Native multimodal support
- Vision encoder support
- Multi-Token Prediction (MTP) support restored
Checkpoint Format
- HuggingFace Transformers
- Safetensors
- FP16
- Sharded checkpoint
Compatibility
Validated with:
- HuggingFace Transformers
- Accelerate
- vLLM
- SGLang
Validation
The reconstructed checkpoint was validated for:
- HuggingFace configuration loading
- Safetensors index integrity
- Tensor mapping consistency
- Sharded checkpoint consistency
- Metadata consistency
- MTP tensor restoration
- Transformers compatibility
Validation summary:
- Tensor count: 1045
- MTP tensors restored
- AutoConfig loading verified
- Safetensors index verified
- HuggingFace checkpoint structure verified
Attribution Notice
This repository is a derivative reconstruction based on publicly released model weights.
The maintainer of this repository:
- Did not create the original model architecture
- Did not perform the original model training
- Did not create the original datasets
- Did not perform the uncensoring process
- Did not create the original GGUF release
All credit for the original work belongs to the respective authors and organizations.
This repository only provides a HuggingFace-compatible reconstruction of publicly available weights.
Known Limitations
Although substantial effort was invested into preserving compatibility and structure, exact numerical equivalence to any original pre-quantization checkpoint cannot be guaranteed.
This repository should be considered a reconstruction of the public GGUF release rather than an official source checkpoint.
Users should independently validate the model for their intended use cases.
Disclaimer
This repository is provided for:
- Research
- Education
- Preservation
- Interoperability
- Compatibility
purposes only.
The maintainer makes no representation regarding:
- Accuracy
- Reliability
- Safety
- Performance
- Suitability
- Fitness for a particular purpose
- Behavioral equivalence
- Numerical equivalence
Users are solely responsible for evaluating the model before deployment or production use.
No Warranty
THE MODEL IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED.
THIS INCLUDES, BUT IS NOT LIMITED TO:
- MERCHANTABILITY
- FITNESS FOR A PARTICULAR PURPOSE
- NON-INFRINGEMENT
- PERFORMANCE
- ACCURACY
- RELIABILITY
- SECURITY
USE OF THIS MODEL IS ENTIRELY AT YOUR OWN RISK.
Limitation of Liability
IN NO EVENT SHALL THE MAINTAINER, CONTRIBUTORS, OR DISTRIBUTORS OF THIS REPOSITORY BE LIABLE FOR ANY CLAIM, DAMAGES, LOSSES, OR OTHER LIABILITY ARISING FROM:
- USE OF THE MODEL
- INABILITY TO USE THE MODEL
- GENERATED OUTPUTS
- MODEL BEHAVIOR
- SECURITY ISSUES
- COMPLIANCE ISSUES
- REGULATORY ISSUES
- DATA LOSS
- BUSINESS INTERRUPTION
- DIRECT OR INDIRECT DAMAGES
USERS ASSUME FULL RESPONSIBILITY FOR ALL CONSEQUENCES ARISING FROM THE USE, DEPLOYMENT, MODIFICATION, OR REDISTRIBUTION OF THIS MODEL.
Content Responsibility
This repository is derived from an uncensored model release.
The maintainer does not endorse, approve, verify, or take responsibility for any content generated by the model.
Users are solely responsible for ensuring compliance with:
- Applicable laws
- Local regulations
- Platform policies
- Organizational policies
- Industry requirements
before deploying, distributing, or using the model.
License
This repository is distributed under the Apache License 2.0.
Please also review the licenses and terms associated with:
- Qwen/Qwen3.6-35B-A3B
- HauhauCS/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive
By downloading, using, modifying, or redistributing this repository, you acknowledge that all use is performed at your own risk.