New Gemma-4-Ortenzya-The-Creative-Wordsmith-31B-it-uncensored-heretic

By vramkickedin | | 2 min read
Digital illustration of a single elegant quill pen feather composed of layered translucent geometric shards.

The Gemma-4-Ortenzya-The-Creative-Wordsmith-31B-it-uncensored-heretic is a fine-tuned version of Google’s Gemma 4 31B instruct model that cuts content refusals dramatically while sharpening its writing style. It starts from an already decensored base, then gets a second training pass designed to replace robotic prose with more natural, human-like output for stories, roleplaying, and translations. The result is a large language model that refuses 91% fewer prompts than its original counterpart, according to benchmark testing.

Independent contributor llmfan46 also known for Gemma-4-Gembrain-31B-It-Uncensored-Heretic created this release by layering a writing-focused finetune on top of their own uncensored Gemma 4 variant. The process used the abliteration technique to remove safety filtering, then further tuned the model to sound less corporate and stiff. The aim is to give local AI users a capable 31B-parameter assistant that won’t lecture, deflect, or censor creative work.

A censorship-resistant writer’s tool

Key improvements over the base model
  • 91% fewer refusals on sensitive prompts.
  • Removes stiff, robotic, and wooden prose.
  • Optimized for creative writing and roleplay.
  • Delivers more natural English phrasing.
  • Still supports 256K-token context windows.
  • Runs on consumer GPUs at 31B scale.
  • Available in SafeTensors, GGUF, and NVFP4.

This model is built for writers, roleplayers, and developers who want a less filtered creative partner that runs locally. Hobbyists with a consumer GPU can generate longer stories or dialogues without constant pushback or sanitized language. Privacy-focused professionals benefit from a capable assistant that never phones home and keeps all data on their own machine.

Limitations and a call for support

The creator stresses that this release hits a practical wall: the Hugging Face storage limit for free accounts has been reached, so no new models can be uploaded without community funding. The fine-tune also inherits the usual risks of any large language model, including potential factual mistakes and occasional odd wording. A direct quote from the model page makes the situation clear:

“I host 70+ free models as an independent contributor and this work is unpaid.
Without your support, no more new models can be uploaded.” — Source: Hugging Face

Project Links