Gemma 4 Abliterated: 26B MoE vs 31B Dense Guide

Gemma 4 abliterated builds are community edits of Google's Gemma 4 that remove its refusals, and for local use the two that matter are the 26B-A4B mixture-of-experts model and the 31B dense model. Google released Gemma 4 on April 2, 2026 under the Apache 2.0 license, and several publishers now offer uncensored versions of both. These builds cut refusals from roughly 99 to 100 out of 100 test prompts down to a handful, while keeping KL divergence low enough to preserve most of the original model's ability.

The two behave differently: the MoE model trades some raw quality for speed, while the dense model is the quality pick. What differs between the builds is the editing tool, the layers changed, and the published numbers. This guide explains what Gemma 4 is, how abliteration works, and how the best-known 26B-A4B, 31B, and E4B builds compare, so you can pick one for your hardware.

What is Gemma 4?#

Gemma 4 is Google's family of open-weight multimodal models, released on April 2, 2026 under the Apache 2.0 license, and it comes in five sizes. For local use, two sizes matter most: the 26B-A4B mixture-of-experts (MoE) model, which runs only 3.8B parameters per token, and the 31B dense model, which uses all of its parameters.

The two differ in design:

• 26B-A4B (MoE): 25.2B total parameters but only 3.8B active per token. It has 30 layers, and selects 8 of its 128 experts plus one shared expert for each token, which keeps generation fast.
• 31B (dense): 30.7B parameters across 60 layers, and it uses every parameter on every token.

Both support up to 256,000 tokens (256K) of context and more than 140 languages, and both add a thinking mode that writes out reasoning before the answer, along with text and image input. At release, the Arena text score was 1452 for the 31B model and 1441 for the 26B-A4B model. See the Google announcement and the Gemma 4 model card for the full figures.

See also: Uncensored vs abliterated: what's the difference?

Why does Gemma 4 refuse, and what is abliteration?#

The instruction-tuned Gemma 4 refuses heavily: in the publishers' refusal tests, the unmodified model declined 99 of 100 prompts, or all 100. Abliteration removes that refusal by editing the weights directly. A 2024 study found that a model's decision to refuse is concentrated along a single direction inside the weights, so removing that direction strips the behavior without retraining.

That method is what makes an uncensored Gemma 4 possible. Abliteration subtracts the refusal direction from the model's weights while leaving its knowledge in place, so it is cheap and fast compared with retraining. The editing tool Heretic automates the process and searches for an edit strength that removes refusals without damaging the model, and its ARA mode is used in most published Gemma 4 builds. For MoE models, some publishers also use EGA, which edits the experts one by one instead of only the dense part.

One number tells you how much a build gave up to get there. KL divergence measures how far the edited model's output has drifted from the original; the smaller it is, the closer the model stays to the base. As a rule of thumb, a build in the 0.05 to 0.12 range has kept most of its ability while removing the large majority of refusals.

See also: What is abliteration?

Which abliterated Gemma 4 builds are available?#

Several, and they are not interchangeable. For 31B there is llmfan46's ARA build; for 26B-A4B there are llmfan46's standard and ultra builds, TrevorJS's EGA build, and SevenOfNine's full-precision build; for the smaller E4B there is InfinimindCreations' build. The table below summarizes the published numbers, and the list that follows gives the details.

31B abliterated builds#

• llmfan46 (gemma-4-31B-it-uncensored-heretic): made with Heretic v1.2.0 and its ARA method, editing layers 30 to 48 at the attention output (attn.o_proj). KL divergence is 0.0541, and refusals fell from 99/100 in the original to 10/100. On a 7,021-question MMLU run it scored 85.90% against the original's 86.50%. GGUF files are available, including Q4_K_M at about 17.4 GB.

26B-A4B abliterated builds#

• llmfan46 heretic (gemma-4-26B-A4B-it-uncensored-heretic): Heretic v1.2.0 with ARA, editing layers 1 to 30. KL divergence is 0.0468, and refusals fell from 100/100 to 11/100. Quantizations run from Q3_K_M to BF16.
• llmfan46 ultra (gemma-4-26B-A4B-it-ultra-uncensored-heretic): the same ARA method pushed further. Refusals fall to 3/100, at the cost of a higher KL divergence of 0.1237.
• TrevorJS (gemma-4-26B-A4B-it-uncensored): norm-preserving biprojection plus EGA (OBLITERATUS) that edits all 128 experts per layer. Refusals fell from 98 to 1 effective refusal on a 100-prompt set, and to 5 out of 686 across four datasets (0.7%). KL divergence is 0.09. The publisher reports that editing only the dense part left 29/100 refusals, and adding EGA brought that to 3/100.
• SevenOfNine (Gemma-4-26B-A4B-It-Abliterated): Heretic run in full bf16 across 200 trials. KL divergence is 0.0845, with 82% of hard refusals removed.

E4B abliterated builds#

• InfinimindCreations (gemma-4-E4B-it-uncensored): biprojection across all 42 layers. On a 656-prompt test it refused 0 times (the original refused 99/100), with a KL divergence of 0.068. GGUF files are available.

26B-A4B MoE or 31B dense: which should you run?#

Choose by VRAM and by what you value. E4B fits around 8 GB and is the cheapest way to try an uncensored build. The 26B-A4B MoE fits around 16 GB and generates fast, because only 3.8B parameters run per token, making it the balanced pick. The 31B dense needs around 24 GB and is the quality pick.

If you are unsure, make the 26B-A4B MoE your first choice when you have about 16 GB of VRAM, and the 31B dense your first choice when you have 24 GB. See also: What is a local LLM?

How do you run an abliterated Gemma 4 locally?#

Download a GGUF file and load it in a runtime such as Ollama or llama.cpp. The 31B model's Q4_K_M file is about 17.4 GB, which with context overhead points to roughly 24 GB of VRAM. The 26B-A4B model's Q4_K_M file is about 16.8 GB, so it does not quite fit in 16 GB on its own; pick a smaller quantization or offload the experts to system memory.

That second option works in practice. SevenOfNine's report runs its 26B-A4B build on an RTX 4080 Super (16 GB VRAM) with 32 GB of RAM, using llama.cpp's -cmoe option to keep the experts in main memory, and reaches 34.5 tokens per second at Q5_K_M. If you do not have a GPU, TrevorJS's build is also served through an inference provider (Featherless), so you can test its behavior before downloading the local files.

See also: GGUF quantization explained

What should you check before using an abliterated Gemma 4?#

What is Deepent Research?#

Deepent Research is a browser-accessible LLM chat service built to refuse as little as possible. It keeps going with research that a standard AI would cut off mid-answer.

• It refuses nothing. We run uncensored and abliterated models with guardrails reduced as far as possible, so prompts a standard AI answers with "I can't help with that" still get a response (Qwen3.5 / Qwen3.8-27B-abliterated, depending on your plan).
• Your input is never used for training. What you look up is used only to provide the service, never to train AI or for any other purpose. Personal data is protected, and our privacy policy is public.
• Try it free. No GPU and no install, just open a browser. In the logged-out trial your conversation isn't saved, so you can start casually.

→ Try it free, no sign-up: https://deepentresearch.com

No question can stop research.

Frequently asked questions#

Is Gemma 4 abliterated an official Google model?#

⌄

No. Google released Gemma 4 as an open-weight model under the Apache 2.0 license, but the abliterated versions are community edits. Each publisher applies its own refusal-removal process, so these builds are not supported by Google and differ from one another.

Should I choose the 26B-A4B MoE or the 31B dense build?#

⌄

Decide by VRAM and speed. Around 16 GB, the 26B-A4B MoE is the better fit, because only 3.8B parameters run per token and generation is fast. Around 24 GB, the 31B dense build is the quality-focused choice, since it uses every parameter on every token.

How much do abliterated Gemma 4 builds differ by publisher?#

⌄

Enough to count as different models. For example, one 26B-A4B build keeps KL divergence near 0.04 to stay close to the original, while another edits the experts and pushes refusals below 1%. The tool, the edited layers, and the published numbers all vary.

Does an abliterated model give accurate answers?#

⌄

Not necessarily. Removing refusals changes whether a model answers, not whether what it says is true. An abliterated Gemma 4 can still produce incorrect statements, so you should verify anything you rely on, especially numbers and factual claims.

Disclaimer#

This article was drafted with AI assistance and then reviewed, fact-checked, and edited by the Deepent Research team before publication.
It is provided for educational and research purposes and is not intended to encourage unlawful activity or crime.
Where figures are cited, they come from third-party public reports and vendor-published values and may vary with measurement conditions.