Skip to content
Computer vision

gemma-4-31B-it-The-DECKARD-HERETIC-UNCENSORED-Thinking

an uncensored version with a reasoning mode

2 min readupdated

Author iBossonline has released a model based on google/gemma-4-31B-it (31.27B parameters) that was first uncensored ("Heretic'ed") and then fine-tuned via Unsloth on a collection of 5 The Deckard datasets. The main goal of the release is to eliminate refusals ("no nanny, no oversight") while boosting reasoning depth, character, and generation quality, outperforming the base version on 6 out of 7 internal benchmarks. It includes two Jinja chat templates: thinking mode (step-by-step reasoning before responding) is hardcoded by default, but it can be switched to a regular instruct mode. The model is a great fit for local enthusiast runs who need a strong, unfiltered multimodal LLM for creative, analytical, and complex multi-step tasks.

Key features

  • Two operating modes: thinking mode is enabled by default via a modified chat_template.jinja, and a chat_template-instruct.jinja file is provided for the classic format.
  • Complete uncensoring: the model underwent a de-censoring pass before fine-tuning and does not block complex or edgy prompts.
  • Fine-tuned on The Deckard: uses a manually gathered and cleaned collection of 5 closed datasets from the "Almost Human" project.
  • Inherited base capabilities: retains support for a 128k token context (and up to 256k per Gemma 4 31B specs), image, video, and code processing, and function calling.

What you can do

  • Write complex creative texts, scripts, and philosophical dialogues between characters with deep logical consistency.
  • Use in local chat clients (with support for <|channel>thought ... blocks) for programming and multi-step analysis.
  • Feed mixed text and image prompts into the model to analyze documents, charts, and interfaces without censorship limits.

Example

The model card provides a generation example (Q4KS quantization with thinking mode enabled) based on a complex prompt about creating 6 plot lines for a "Black Mirror" episode focusing on time travel and romance. Before outputting the final scripts, the model independently breaks the task down into requirements within a <|channel>thought block, generates 6 concepts, outlines a three-act structure, characters, and themes for each, and then checks compliance with the show's tone.

Read moreDetails
Internal benchmarks (qx86-hi, Instruct mode vs base): ARC 0.573 (base 0.496), ARC-e 0.719 (base 0.653), HellaSwag 0.715 (base 0.624), OBQA 0.420 (base 0.380), PIQA 0.781 (base 0.732), WinoGrande 0.695 (base 0.653), BoolQ 0.900 (base 0.901).
In Thinking mode, scores on these tests are lower than in Instruct (e.g., ARC 0.431, HellaSwag 0.670, PIQA 0.766).
Context: 128k tokens; proper rendering of the reasoning block in the UI may require configuring the <|channel>thought and tags.
Discuss Telegram 0 18

Will it run?

Runs in 4-bit

Full precision does not fit, a quantised version does: quality barely changes down to 8 bits, more noticeably at 4 bits.

  • BF16weights 58 GB · ready files65 GB
  • FP8 / INT8recommendedweights 29 GB33 GB
  • 4-bitweights 16 GB19 GB

Tip: go 8-bit.FP8 / INT8 is within about a percent of full precision in quality, takes half the memory and runs faster on RTX 40/50 and H100. Full precision only pays off for fine-tuning.

Estimate: weights plus working memory. Long videos, high resolutions and long contexts need more.

Run it on Massed Compute

The script downloads the model to the rented machine. How to run it is in the model card.

  1. 1. Rent a machine with a fitting GPU

    Open Massed Compute
  2. 2. Paste the script into the machine's terminal

    Connect over SSH (the Massed Compute panel shows the command) or open a terminal on the machine's desktop.

    HF_TOKEN / CIVITAI_TOKENThe script starts with token lines: paste yours between the quotes. The Hugging Face token is for gated models, the Civitai one for Civitai models.

    # iBossonline/gemma-4-31B-it-The-custom-thinking
    mkdir -p ~/models
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv tool install "huggingface_hub[cli]"
    hf download iBossonline/gemma-4-31B-it-The-custom-thinking --local-dir ~/models/gemma-4-31B-it-The-custom-thinking
    # The model is in ~/models. How to run it is in the model card (link above).
  3. 3. Use it

    The model is in ~/models. How to run it is in the model card (link above).

Stop the machine in the panel when you are done: it bills by the minute while it runs.

How to speed it up

Picked for this model type. Choose your hardware above to see what fits you exactly.

  • FlashAttention 44.0.0b40from November 18, 2026new release

    fastest exact attention on B200

    Kernels for Blackwell (beta). Exact attention without quantisation — fine for everything, LLMs included.

    How to enable: pip install flash-attn-4; vLLM and SGLang pick it up themselves.

    Earlier generations (2)
    • FlashAttention 3 — The Hopper (H100/H200) version with FP8.
    • FlashAttention 2 — The standard for RTX 30/40/50 and A100: exact attention without the memory cost of long contexts.
  • torch.compile2.15.0from October 28, 2026new release

    10–30% for free

    Compiles the model for your GPU. Slower first run, faster afterwards.

    How to enable: model = torch.compile(model); in ComfyUI the TorchCompileModel node.

Model family

  1. gemma-4-31Bfamily root
  2. gemma-4-31B-it
  3. gemma-4-31B-it-The-custom-thinking

New posts about versions of gemma-4-31B (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 12All on Hugging Face
Fine-tunes 12All on Hugging Face
Merges 3All on Hugging Face

About the model

iBossonline/gemma-4-31B-it-The-custom-thinking
Author
iBossonline
Category
Computer vision
Released
Posted

On NeyroLearn

Hugging Face

180

Ask AI about this model

Answered by Gemini 4 · the chat is kept only in this tab

Comments