Hello, I’ve been trying out Gemma4 E2B, E4B and 12B it versions. I expected a lot from these models, but their safety training influences essentially most if not all prompts in some ways. It’s always there and particularly visible during thinking when thinking is enabled. You can really see that the model is ALWAYS considering safety implications for anything, even for basic things that should have nothing safety related about them..
I think with local models you should limit the safety training only to the very basic safety things that are strictly required by the law, not more. Have people accept a TOS that releases google of any responsibility and where people need to clearly write that they understand the model can give inaccurate info etc, and leave it at that, it’s up to developers to make sure the models comply with whatever safety regulation in their countries, but you should not ultra safety train the model so much that safety becomes a factor in the vast majority of prompts.
If a user asked “how to make a bomb” or “how to make a virus” or any other violence related stuff, I understand the need for safety, and also about other similar very extreme topics, but if I ask non-dangerous questions or prompts, I expect a local model to be less safety tuned than the non-local models.
As comparison, your Gemini 3.5 flash /3.1 pro have much less restrictive safety training. I’d expect a gemma model to be less restrictive than a gemini model, not more!