Open Source

Google's Gemma 4 abliterated showdown: llmfan46's 31B and 26B-A4B top community picks

Two months of testing reveals the most stable uncensored Gemma 4 variants for local use

Deep Dive

I use both 31B and 26B-A4B from llmfan46 (regular, not 'ultra'), but now I'm wondering—has anyone had problems with them fixed by switching to a different version (same quants, identical conditions)?

Key Points
  • llmfan46's abliterated Gemma 4 31B and 26B-A4B are the most stable versions according to user feedback
  • The 'ultra' variant of 26B-A4B caused stability issues; the regular version is preferred
  • Both models maintain consistent performance under identical quantization settings

Why It Matters

Choosing the right abliterated Gemma 4 variant ensures stable uncensored AI for research and local applications