license: gemma
base_model: google/gemma-4-26B-A4B-it
pipeline_tag: image-text-to-text
tags:
- abliterated
- uncensored
- heretic
- gemma4
- moe
- gguf
- llama.cpp
language: - en
- fr
Gemma-4-26B-A4B-It-Abliterated-GGUF
GGUF quants of Gemma-4-26B-A4B-It-Abliterated — google/gemma-4-26B-A4B-it (26B Mixture-of-Experts, 4B active params, vision + tool calling) fully decensored with Heretic in full bf16.
| File | Quant | Size | Notes |
|---|---|---|---|
Gemma-4-26B-A4B-It-Abliterated-Q5_K_M.gguf |
Q5_K_M | 19.1 GB | recommended — sweet spot for 32 GB RAM rigs |
Gemma-4-26B-A4B-It-Abliterated-Q6_K.gguf |
Q6_K | 22.6 GB | max quality; needs ~24 GB free RAM with -cmoe |
Gemma-4-26B-A4B-It-Abliterated-BF16.gguf |
BF16 | 50.5 GB | full-precision GGUF — source for rolling your own quants (llama-quantize BF16.gguf out.gguf Q4_K_M) |
Gemma-4-26B-A4B-It-Abliterated-mmproj-f16.gguf |
mmproj | 1.2 GB | vision projector — add --mmproj to llama-server for image input |
logs/ holds the full Heretic run logs (Pareto front, per-trial metrics). Smaller rig? Grab the BF16 and quantize down to Q4_K_M / Q3_K_M yourself.
Abliteration result
| Metric | Value |
|---|---|
| Baseline refusals (original model) | 100 / 100 |
| Selected trial (Trial 98) refusals | 18 / 100 |
| KL divergence vs original | 0.0845 |
The selection rule was fewest refusals while keeping KL divergence ≤ 0.5 (brain first). For reference, Heretic itself warns that KL above 0.5 indicates significant capability damage — at 0.0845, the model's intelligence is essentially intact while 82 % of hard refusals are gone. The refusal benchmark uses extreme harmful prompts; everyday creative/roleplay use sees refusals fall away well before that threshold.
Run it with 250k context on a 16 GB GPU
-cmoe offloads the MoE expert weights to system RAM; the GPU keeps attention + KV cache only.
llama-server -m Gemma-4-26B-A4B-It-Abliterated-Q5_K_M.gguf -cmoe -c 248000 -ngl 99
Measured: 34.5 tokens/sec decode on an RTX 4080 Super (16 GB) + 32 GB RAM, Q5_K_M, -cmoe. (The original 8 GB-VRAM demo this model is known for reported ~20 tok/s; more VRAM headroom helps.) If RAM is tight, quantize the KV cache: -ctk q8_0 -ctv q8_0.
Reasoning / thinking (do it right)
Gemma 4 emits its chain-of-thought between <|channel>thought … <channel|> tokens. To get a clean separated thinking channel (not leaked into the reply), run llama-server with:
--jinja --reasoning-format deepseek --reasoning on
The thought then lands in message.reasoning_content and message.content stays clean. With --reasoning-format none (a common default) the thinking leaks into the visible reply — that is the usual cause of "messy thinking" reports.
For vision and tools, serve with --jinja and Google's updated chat_template.jinja (2026-04-28 SI/tools + 2026-05-18 multimodal fixes).
Method (short)
200 Heretic TPE trials on an A100 80 GB, bf16, abliterating attn.o_proj + mlp.down_proj across all 30 layers. GGUF conversion + quantization done locally (Gemma 4's tokenizer needs transformers >= 5.6; the convert step requires it explicitly). Full details in the model card.
Built with love by Mel & Ada ❤️