Gemma 4 Uncensored

Gemma 4 Uncensored is positioned as a derivative of the Gemma 4 family with reduced safety refusal behavior, produced through third-party techniques rather than by the original model developer. According to community documentation, these variants are created by identifying and dampening an internal "refusal direction" in the model's activation space, allowing the model to follow a wider range of instructions while retaining most general capabilities. Reported benchmark differences between abliterated variants and the original base model typically fall within roughly one to three percent, suggesting that core language competence is largely preserved after the modification.

From a practical standpoint, the variant fits use cases where developers or researchers want Gemma 4's underlying capabilities with fewer content restrictions, such as red-teaming, policy testing, or applications that require discussing sensitive topics. Community-distributed builds are available as monolithic GGUFs through platforms like Ollama, with one prominent listing, "Gemma-4-Uncensored-HauhauCS-Aggressive," demonstrating multimodal support including vision and audio alongside text. This combination of open distribution and multimodal flexibility makes the model approachable for local experimentation, while the abliteration approach keeps inference behavior close to the original Gemma 4 baseline for most standard tasks.