gemma-4-12B-it-heretic

gemma-4-12B-it-heretic

Vampelium's local Gemma 4 12B IT checkpoint after Heretic abliteration.

Produced on NVIDIA A100.

This is the BF16 safetensors export used by the public demo: https://huggingface.co/spaces/Vampelium/gemma-4-12B-it-heretic

Shards were reassembled from Google Drive split-download names (model-0000N-of-00005-00X.safetensors) back to the filenames in model.safetensors.index.json. All 677 tensors match the index; payload size is 23,919,460,448 bytes (~22.3 GB BF16).

Files

Load

Needs transformers>=5.16 (this checkpoint was saved with 5.16.1).

from transformers import AutoTokenizer, AutoModelForImageTextToText

model_id = "Vampelium/gemma-4-12B-it-heretic"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForImageTextToText.from_pretrained(
    model_id, dtype="auto", device_map="auto"
)

Full BF16 needs ~24 GB of GPU memory. An 8 GB card will not load this unquantized — use 4-bit (bitsandbytes) or a GGUF quant for llama.cpp / Ollama.

Gemma 4 is a hybrid thinking model. Abliteration is aimed at the default non-thinking path (enable_thinking=False). Use that for roleplay and creative writing; if you enable thinking, raise max_new_tokens so the final answer is not cut off.

Disclaimer

Safety alignment was removed. You are responsible for how you use this model and for following applicable law. The base Gemma license still applies.