Gemma 4 users debate uncensored variants

This is less a launch than a temperature check on the post-release Gemma 4 ecosystem: the community is rapidly sorting models by practical refusal behavior, roleplay quality, quantization, and Apple Silicon support.

• –Base Gemma 4 may be "uncensored enough" for many local users, which weakens the need for aggressive safety-removal fine-tunes.
• –The uncensored ecosystem is fragmenting by use case: EZCon for smaller MLX builds, DavidAU/Heretic for broad decensoring, and TheDrummer-style tunes for character and roleplay quality.
• –Developers should treat these variants as workload-specific forks, not drop-in upgrades; decensoring can change instruction following, benchmark behavior, and reliability.
• –The thread reinforces how fast open-weight model communities move after release: within weeks, users are already comparing abliterated, quantized, and fine-tuned derivatives.