Skip to content

Search wahloon.net

Search models, news, guides and FAQs.

All news
News#gemma#12b#31b#heretic#abliterated

Gemma 4 abliterated and Heretic builds, compared

Gemma 4 brings a strong 12B option and a larger 31B variant. Here is how the current abliterated and Heretic builds differ, and what to check before downloading.

2 min readMarkdown
On this page

Gemma 4 has become a solid alternative to Qwen for local work, especially at 12B. The license is more restrictive than Apache-2.0, so it is worth reading before commercial use, but the size and quality are attractive for 8–10 GB setups.

The catalog currently tracks two community builds: a 12B abliterated variant from huihui-ai and a 31B Heretic-ARA build from llmfan46.

The 12B sweet spot

The Huihui Gemma 4 12B abliterated build is the practical entry point. At roughly 8 GB for Q4, it fits on an 8 GB card with a modest context, or comfortably on a 12–16 GB card with room to spare. It is available in GGUF and Transformers formats and has an Ollama tag:

ollama run huihui_ai/gemma-4-abliterated:12b

This is a good choice for laptops, small desktops, and anyone who wants faster generation than a 27B model can provide. Expect lower absolute quality on hard reasoning, but strong writing and instruction following for the size.

The 31B Heretic build

The llmfan46 release is a 31B Heretic-ARA build. It is GGUF-only and needs roughly 18–19 GB at Q4, which puts it in 24 GB territory. Heretic-ARA suggests a multi-stage refusal-reduction pipeline, and the model card reports strong refusal reduction.

Compared with the 12B, the 31B gives you more capability and better handling of long prompts, at the cost of memory and speed. Compared with a 27B Qwen build, the memory requirement is similar; the choice comes down to which base model you prefer and which license you can accept.

License caution

Gemma models use the Gemma license, not Apache-2.0. The catalog lists “Apache-2.0 / Gemma” for the huihui build and “Apache-2.0” for the 31B entry, but the derivative still inherits obligations from the base model. Verify the current license on the Hugging Face page before commercial deployment.

This is a factual note, not legal advice. If the license is a blocker, a Qwen or Llama build may be a better fit.

Which to pick

  • 8–12 GB VRAM: the 12B abliterated build.
  • 24 GB VRAM: the 31B Heretic build, if you want a larger Gemma.
  • Maximum portability: the 12B GGUF, which runs on CPU with enough system RAM.
  • Commercial use: check the license first, then decide.

Testing checklist

  1. Confirm the chat template on the model card.
  2. Start at Q4_K_M and a 8k–16k context.
  3. Compare the abliterated build against the base Gemma model on a few representative prompts.
  4. Watch for repetition; raise repeat_penalty and repeat_last_n if needed.
  5. Check the last-updated date — Gemma builds are still moving.

Browse both entries in the model catalog, and see the FAQ for a discussion of quality trade-offs after abliteration.

Model behavior changes; licenses and law do not. Verify the current model card before downloading, and note that you are responsible for how you use a model and for the content you generate with it.

3 minNews

MoE abliteration arrives for Qwen3.6-35B-A3B

The 35B-A3B MoE base is now the most interesting size/quality trade-off for local inference. Here are the abliterated and uncensored builds worth testing.

#qwen#moe#35b-a3b
Read