Launch Announcement: Gemma 4 Uncensored Models (E4B & E2B) – Fully Multimodal, Aggressive Variants
We are excited to announce the release of the first uncensored versions of the Gemma 4 series, featuring two new models: the Gemma 4 E4B (4 billion parameters) and the Gemma 4 E2B (2 billion parameters). Both models are tailored to deliver an aggressive, uncensored output—meaning they have no refusal mechanism—directly reflecting the original Google release, free from modifications or personality adjustments.
Key Highlights
- Uncensored & Aggressive: These models provide direct, unfiltered responses, with zero refusal instances (0/465 refusals noted on both models). They are intended for users seeking the authentic capabilities of the original architecture without any moderation layers.
- Fully Multimodal: Designed to handle multiple input types seamlessly—text, images, videos, and audio—all within a single model. This makes them highly versatile for various complex applications.
- Compatibility & Support: The models include a
.mmprojfile that supports vision and audio inputs. They are compatible with popular frameworks such as llama.cpp and LM Studio, and are designed to work with any tool capable of reading GGUF format.
Model Details & Download Links
Gemma 4 E4B (4B parameters):
Download from Hugging Face
Gemma 4 E2B (2B parameters):
Download from Hugging Face
Both models feature extensive quantization options—such as Q8_K_P, Q6_K_P, and Q5_K_P—that optimize performance and size, generated using the imatrix tool. Notably, the K_P quantizations utilize model-specific analysis to preserve quality, achieving levels only 1–2 quantization steps lower at a modest increase of 5–15% in file size.
Technical Specifications
| Feature | Gemma 4 E4B | Gemma 4 E2B |
|—|—|—|
| Number of Layers | 42 | 35 |
| Context Capacity | 131K tokens | 131K tokens |
| Attention | Mixed sliding window + full attention | Mixed sliding window + full attention |
| Multimodal Support | Yes | Yes |
| Memory Optimization | KV shared layers | KV shared layers |
Sampling Settings:
– Temperature: 1.0
– Top-p: 0.95
– Top-k: 64
(Use the --jinja flag with llama.cpp to enable these settings.)
Note: Hugging Face’s hardware compatibility widget may not recognize K_P quantizations; select “View +X variants” or browse the “Files and versions” tab to access all download options. The model may show “?” for K_P in LM Studio; this is purely cosmetic, and the models load correctly.
Future Releases & Development Roadmap
The upcoming models in the pipeline include:
- Gemma 4 E31B (Dense variant)
- Gemma 4 E26B-A4B (Mixture of Experts – MoE)
Development efforts are currently focused on refining these larger models, ensuring they meet quality standards before release.
Additional Remarks
The models leverage advanced techniques similar to NVIDIA’s GenRM, incorporating generative reward mechanisms that act as internal critics. While this enhances uncensoring capabilities, it also makes manual testing more complex, especially over longer contexts. Users are advised to set realistic expectations—these models aim for broad applicability but are not designed to surpass larger, more resource-intensive models above 7B parameters.
Resources
All models are accessible via Hugging Face – HauhauCS. Stay tuned for future updates and breakthroughs.
Note: As part of ongoing projects, we are working on innovative developments that will be announced soon—stay connected for the latest updates.
This release marks a significant step forward in providing uncensored, multimodal AI models that faithfully mirror Google’s original models, empowering users with versatile and unfiltered AI capabilities.
