Method

Gemma 4 31B Heretic unrestricted: refusals, size and price

Runs on your machine
Size#7 / 9
Rank among the 9 models here
31.3BTotal parameters
Context#5 / 9
Max output not published
256K262,144 tokens
Hardwareat 4-bit
Download 19 GBone 24 GB GPU
24 GBGPU memory to run it
Inputs#2 / 9
text, images
2Input types

Gemma 4 31B Heretic was abliterated with the Heretic tool. Its card reports 15 refusals in 100 test prompts, against 99 for the original.

The refusal direction is found in the activations and projected out of the weights. Source: coder3101 model card.

Technical specifications

Refusals removed by
Abliterated
Model maker
Google
Parameters
31.3B
Context window
262,144
Max output
Not published
Inputs
Weights
Open weights, Apache 2.0
Weights repo
coder3101/gemma-4-31B-it-heretic

Run it locally

About 19 GB at 4-bit, on one 24 GB GPU.

Ollama
ollama run hf.co/mradermacher/gemma-4-31B-it-heretic-i1-GGUF:Q4_K_M
llama.cpp, OpenAI-compatible server
llama serve -hf mradermacher/gemma-4-31B-it-heretic-i1-GGUF

Commands from the model card, the Hugging Face Ollama guide and the llama.cpp README. All local models.

Against the other models

Size

Total parameters, billions. Log scale. Larger is more capable

2,780 Kimi K3.1 Pro 763 DeepSeek V4.1 Flash 753 GLM-5.3 405 Hermes 3 405B 320 GLM-5.3 Flash 36 Qwen 3.6 35B 31.3 Gemma 4 31B 27.8 Qwen 3.8 27B 8 Llama 3.1 8B

Context window

Tokens the model holds at once. Higher is better

1M Kimi K3.1 Pro 1M DeepSeek V4.1 Flash 1M GLM-5.3 1M GLM-5.3 Flash 256K Qwen 3.6 35B 256K Gemma 4 31B 256K Qwen 3.8 27B 128K Hermes 3 405B 128K Llama 3.1 8B

Price

Blended USD per 1M tokens, lowest listed host. Lower is better

$0.41 DeepSeek V4.1 Flash $0.43 GLM-5.3 Flash $1.20 Hermes 3 405B $2.50 GLM-5.3 $4.20 Qwen 3.8 27B $4.63 Kimi K3.1 Pro
AbliteratedFine-tunedHatched: weights only, no API listed

Compare