Modelpublic

deucebucket/Gemma-4-26B-A4B-it-Heretic-Cerebellum-GGUF

sourceHugging Facegemmaupdated 3mo agoView on Hugging Face

7likes500downloads

Model Card

Gemma 4 26B-A4B-it Heretic Cerebellum GGUF

Sensitivity-guided mixed-precision GGUF of coder3101/gemma-4-26B-A4B-it-heretic, a decensored variant of google/gemma-4-26B-A4B-it. It uses the Cerebellum v6 tensor allocation transferred verbatim onto the heretic weights. The shipped file carries Google's updated Gemma 4 chat-template metadata (2026-05-18 state) with zero tensor changes. Versions appear in filenames, not the repo name.

Files

File	Description
`gemma-4-26B-A4B-it-heretic-cerebellum-v1.1-templatefix-Q3_K_M.gguf`	~11 GB; v1 allocation + updated chat-template metadata
`gemma-4-26B-A4B-it-heretic.mmproj-f16.gguf`	vision projector (required for image/video)

Evaluation

Measured directly on the GGUF with llama.cpp llama-server on an RTX 3090, temperature 0, project benchmark harness. The v1.1 templatefix file is metadata-only over v1, so these describe the same weights. The comparison column is our own plain (non-heretic) Cerebellum v6 build on the same harness, shown so the abliteration cost is visible. Summary JSONs and per-question samples are in benchmark_results/.

Benchmark	Heretic v6 alloc (11 GB)	Plain Cerebellum v6 (11 GB)
ARC-Challenge (1172 q)	95.48%	95.56%
HellaSwag (10042 q)	83.49%	84.55%
MMLU-Redux (2400 q)	71.42%	71.33%
HumanEval base (chat, no-think)	92.07%	pending re-audit
HumanEval+ (chat, no-think)	89.63%	pending re-audit
Vision smoke	6/6	—

HumanEval/HumanEval+ used the chat-completions harness (scripts/benchmark_evalplus_chat.py, enable_thinking: false, thinking_budget_tokens: 0, BENCH_WORKERS=1, max_tokens: 768). The completion audit for that run recorded 0 prompt echoes, 0 repeated function definitions, 0 pass-only outputs, and 2 genuine syntax failures. The plain v6 HumanEval artifacts were raw-completions and are marked for re-audit, so no plain-v6 HumanEval is published.

Usage

Gemma 4 requires --jinja. For non-thinking output, pass request-level chat_template_kwargs: {"enable_thinking": false} and thinking_budget_tokens: 0; do not set a fixed server --reasoning-budget (it can burn output into hidden reasoning until the length cap, which looks like a repetition loop).

bash

llama-server \
  --model gemma-4-26B-A4B-it-heretic-cerebellum-v1.1-templatefix-Q3_K_M.gguf \
  --mmproj gemma-4-26B-A4B-it-heretic.mmproj-f16.gguf \
  -ngl 99 --ctx-size 65536 --parallel 1 --flash-attn on \
  --cache-type-k q8_0 --cache-type-v q8_0 --jinja --reasoning auto

Measured on one RTX 3090 (24 GB), KV q8_0: context to 131,072. This rig's measurements; no quality claims beyond them.

Provenance

—Source (heretic): coder3101/gemma-4-26B-A4B-it-heretic — abliterated variant
—Original family: google/gemma-4-26B-A4B-it — Google Gemma Team
—Recipe: Cerebellum v6 tensor allocation transferred to the matching heretic layout; v1.1 is a chat-template metadata refresh (Google 2026-05-18 template), zero tensor changes

Credits

—Source model: coder3101/gemma-4-26B-A4B-it-heretic
—Original Gemma family: Google Gemma Team
—GGUF runtime: llama.cpp
—Quantization method: Cerebellum — deucebucket