Announcing the Release of Gemma4-26B-A4B Uncensored Balanced Model with K_P Quants
We are thrilled to share a significant milestone: surpassing 10 million downloads on Hugging Face! This number reflects the popularity of our models, exclusively counted from unique downloads on my personal account, excluding duplicates, fine-tunes, and other variations.
After over a month of dedicated development, I am pleased to announce the release of the Gemma4-26B-A4B Uncensored Balanced model candidate (Release Candidate). This version represents the most challenging uncensored model I’ve built to date, aiming for robust performance while maintaining safety and stability.
Highlights of the Gemma4-26B-A4B Uncensored Balanced Model
- Performance and Safety: Achieved a perfect record of 0 refusals across 465 tested prompts, demonstrating my goal of near-zero automatic refusals in typical use cases.
- Design Philosophy: The ‘Balanced’ setting offers measured reasoning on provocative or edgy topics, providing thorough but unfiltered responses. It inserts brief safety cues when handling sensitive content, ensuring output remains comprehensive and faithful to the original prompt.
- Implementation Details: This release is a fully uncensored variant of the original Gemma4-26B-A4B. The upcoming ‘Aggressive’ version will strip even the safety framing for direct, no-narrative responses—a work in progress, as I refine its lossless output quality.
Model Capabilities and Use Cases
The Balanced variant is optimized for scenarios such as creative writing, role-playing, and emotional intelligence applications. It reasons through complex or controversial requests, attaching short safety frames before delivering answers, making it the ideal default for most users.
Note: While Gemma4-26B-A4B is excellent for these tasks, for agentic coding, tool use, and related functions, I recommend the Qwen3.6 model, which consistently outperform in these domains.
Technical Specifications
- Model Size: 25.2 billion parameters, with 3.8 billion active parameters utilizing a mixture of experts (MoE) architecture.
- Architecture:
- 30 layers with hybrid attention mechanisms, including sliding window and full global attention.
- Hidden size: 2,816; Heads: 16; KV heads: 8.
- Native context length: 262,144 tokens.
- Multimodal Support: Compatible with both text and image inputs via combined models and the mmproj tool.
- Quantization and Compatibility:
- Utilizes custom K_P quants, offering approximately 1-2 quality levels uplift with a 5-15% increase in file size.
- Fully compatible with llama.cpp, LM Studio, and any platform supporting GGUF format.
- Please note: some interfaces like HF’s widget or LM Studio may display K_P quants as “?”, but the models function correctly.
Sampling Parameters and Deployment Tips
For optimal results, use the following sample generation settings as recommended by Google:
- Temperature: 1.0
- Top-p (nucleus sampling): 0.95
- Top-k: 64
Additional notes:
– Enable the –jinja flag with llama.cpp for prompt templating.
– Position image inputs before text prompts when using vision features.
– The model performs reliably across re-runs, maintaining long-context coherence without looping issues.
Access and Support
All models, including facets like vision support, are available at our Hugging Face repository: HauhauCS Models.
For updates, community discussions, and to stay informed on upcoming projects, join our Discord community through the links provided in the Hugging Face repository.
Final Remarks
This release embodies months of meticulous work and testing, with extensive automated and manual benchmarks confirming the near-zero refusal rate in typical scenarios. Feedback from users and community members is invaluable, and I plan to continue refining these models to better serve diverse use cases.
Thank you for your continued support and enthusiasm as we push the boundaries of open, uncensored AI models. Enjoy exploring the new Gemma4-26B-A4B Uncensored Balanced model!
