Introducing the Qwen 3.6-35B-A3B Uncensored Aggressive Model: Cutting-Edge AI with K_P Quants Support
We are excited to announce the release of the latest iteration of the Qwen series: the Qwen 3.6-35B-A3B Uncensored Aggressive model. Building upon the foundation of previous models, this version offers significant advancements in performance, versatility, and user customization.
Overview of the Qwen 3.6-35B-A3B Aggressive Model
The new model is based on the Qwen 3.6 architecture, featuring a 35-billion-parameter mixture of experts (MoE) framework. It retains the same MoE size as the earlier 3.5-35B release but benefits from the improvements inherent to the 3.6 base. Notably, this variant is designed to operate without any content refusals, ensuring an “uncensored” and unrestricted output—hence the designation “Aggressive.” Importantly, this version introduces no personality modifications or behavioral changes, maintaining the core qualities of the original Qwen release.
Access and Deployment
The model can be accessed via Hugging Face here: Qwen 3.6-35B-A3B Uncensored Aggressive. It is fully unlocked, with no capability losses—delivering a seamless experience for developers and researchers alike. Users are encouraged to explore its capabilities, noting that the model is free from refusals and operates efficiently with no reported issues such as looping or degradation during extensive testing.
Customization Options
To tailor the model’s behavior, especially regarding its “thinking” process, users can edit the Jinja templates or utilize keyword arguments such as {"enable_thinking": false}. This flexibility allows for optimized performance aligned with specific application needs.
Included Features and Support
The release incorporates a broad array of optimized quantizations, including:
- Q8_K_P
- Q6_K_P
- Q5_K_P
- Q4_K_P
- Q4_K_M
- IQ4_NL
- IQ4_XS
- Q3_K_P
- IQ3_M
- Q2_K_P
- IQ2_M
Additionally, the model supports multimodal functionalities, integrating text, images, and video inputs. Vision support is enabled via the mmproj module, allowing for richer multimedia interactions. All quantizations are generated using the imatrix tool to ensure optimal performance.
The K_P Quants Advantage
For those who missed the previous 122B release, it’s worth noting that K_P quants are a range of custom quantization profiles designed to preserve model quality by employing model-specific analysis. Each model receives an optimized profile, resulting in approximately 1-2 levels of quantization quality improvement at a modest 5-15% increase in file size. These quantizations are compatible with popular frameworks such as llama.cpp and LM Studio, and can be used with any software that supports GGUF format. Please note that recognition of K_P quants may vary in some interfaces like LM Studio, where they might display as “?”, but functionality remains unaffected.
Technical Specifications
- Total Parameters: 35 billion (with approximately 3 billion active)
- Expert Routing: 256 experts with 8 routed per token
- Context Length: 262,000 tokens
- Multimodal Capabilities: Supports text, image, and video inputs
- Attention Mechanism: Hybrid approach combining linear and softmax attention (ratio approximately 3:1)
- Model Architecture: 40 layers
Sampling and Usage Guidelines
During testing, the following sampling parameters were effective:
- Temperature: 1.0
- Top-k sampling: 20
- Repetition Penalty: 1
- Presence Penalty: 1.5
- Top-p (nucleus sampling): 0.95
- Minimum probability: 0
For optimal results, users are encouraged to consult the official Qwen recommendations, which suggest varying settings based on whether “thinking” mode is enabled or disabled. When deploying with llama.cpp, ensure to include the --jinja flag; note that K_P quants may appear as “?” in certain interfaces like LM Studio, but this is purely cosmetic.
Additional Resources and Community
All models, including this latest release, are available on Hugging Face at HuggingFace-HauhauCS. To foster collaboration and support, a dedicated Discord community has been established—links are available in the Hugging Face repository. The community is active, sharing updates, project ideas, and troubleshooting assistance.
Conclusion
The Qwen 3.6-35B-A3B Uncensored Aggressive model represents a significant step forward in open-source AI development. Its uncensored nature, extensive multi-modal support, and optimized quantization profiles make it a versatile tool for researchers, developers, and enthusiasts seeking the highest levels of performance and flexibility.
Feel free to explore, experiment, and join the community to stay updated on future developments.
Note: This article aims to provide a comprehensive and professional overview suitable for publication on a WordPress blog, ensuring clarity and technical accuracy.
