Gemma 4 "Cracked" Status: Which Model Was Modified, What Changed, and What Risk You Take

As of April 6, 2026, the community “cracked/uncensored” discussion is mainly about derivatives of google/gemma-4-31B-it.

This post is only about the concrete facts you care about:

1) Which Gemma 4 Model Has Been "Cracked"?

The main target in public model cards is:

The biggest wave is around this 31B instruct checkpoint and its derivatives.

The commonly claimed change is:

In several model cards, maintainers explicitly describe refusal-removal methods and provide before/after refusal metrics.

Reported by maintainers:

Lower refusal rate on previously blocked prompts
Higher “compliance” to instructions
Similar general capabilities retained for normal tasks (their claim, not guaranteed)

In plain terms: it answers more prompts that base instruct versions would reject.

Examples you can directly verify right now:

Even if the model is publicly downloadable, your risk does not disappear:

Gemma derivatives are still under Gemma terms constraints, including prohibited-use rules and distribution obligations.

If your product produces illegal, infringing, or harmful output, responsibility stays with deployers/operators.

Using high-risk uncensored variants can fail enterprise review, ad platform policy checks, or app-store compliance.

Lower refusal increases abuse surface for jailbreak-style prompts and malicious user attempts unless you add your own guardrails.

For the current “Gemma 4 cracked” narrative, the core fact is simple:

The community mostly modified google/gemma-4-31B-it derivatives to reduce refusal behavior.
Multiple such variants are already publicly listed on Hugging Face.
Technical availability does not remove license, legal, or product liability risk.