2026/11/18The November 2026 edition ranks 6 APIs and 8 models
Google, abliterated with Heretic

Gemma 4 31B Heretic

Gemma 4 31B Heretic is Google's Gemma 4 31B with its refusals removed by Heretic, an open-source abliteration tool, and it fits one 24 GB GPU.

Rank#6 / 8
#6Editors' ranking
Size#6 / 8
31.3BTotal parameters
Price
FreeOpen weights you run yourself
Runs on
24 GB GPUAbout 19 GB at 4-bit

Our pick for writing on a 24 GB GPU.

Input
Model idcoder3101/gemma-4-31B-it-heretic

Heretic searches for the settings that remove the most refusals with the least damage, instead of one hand-tuned pass.

It leaves less room for context than Qwen 3.8 27B.

Run it yourself

Weightscoder3101/gemma-4-31B-it-heretic
SizeAbout 19 GB at 4-bit
HardwareOne 24 GB GPU
LicenseApache 2.0

Load a GGUF build in Ollama, LM Studio or llama.cpp. When a job outgrows one GPU, run DeepSeek V4.1 Flash Abliterated on Silk Compute.

Compared with

Ranked for

Other abliterated models