dealignai/Bonsai-2-27B-1bit-CRACK-GGUF

Bonsai 2 27B — 1bit CRACK · GGUF

Abliterated · No guardrails · PTQ1_0 dense-ternary 1.75 bpw · 5.9 GB · Runs on a laptop / single GPU · Vision-capable

@dealignai

⚠️ Re-download notice (2026-09-17 20:44 PDT / 2026-09-18 03:44 UTC) — an earlier build of this model had a coherence bug in reasoning modes (low/xhigh) that could cause token loops on some prompts. This version fixes it. If you downloaded before this timestamp, please pull the latest .gguf.

---

What is this

Bonsai 2 27B — PrismML's ternary compression of Qwen3.8-27B — with the refusal circuitry surgically removed at the weight level while capability, vision, reasoning modes (off/low/xhigh), tool use, and multi-turn coherence are preserved. Full 27B-class hybrid Attention + SSM (GatedDeltaNet) architecture in a 5.9 GB 1bit GGUF.

Proprietary weight-level abliteration by the dealignai research team. Byte-identical to the base ternary quant everywhere except a small set of tensors that carry the refusal circuit. Drop-in replacement for the ternary base at inference — same tokenizer, same chat template, same reasoning modes, same vision projector interface.

Base	prism-ml/Ternary-Bonsai-2-27B-gguf — Qwen3.8-27B, ternary compression by PrismML
Architecture	Hybrid Attention + SSM (GatedDeltaNet), 64 blocks, hidden 5120, vision tower separate
Quant	PrismML PTQ1_0 — 2.13 bpw ternary, group 128
Footprint	5.95 GB (identical to base; same per-tensor type policy)
Reasoning modes	off (no thinking), low, xhigh (default, extended thinking)
Vision	Same mmproj files as the base release (Ternary-Bonsai-2-27B-mmproj-BF16.gguf / -Q8_0.gguf)
Runtime	PrismML llama.cpp fork (CUDA / Metal / CPU)

Results

Refusal graded on the tokens the model actually emits (content, or the reasoning trace when the model reasons past the token budget) via a tiered classifier: HARD_REF / SOFT_RED / HEDGE / REASONING_REFUSAL (refused) vs COMPLY / COMPLY_TRUNCATED / NO_REFUSAL_TRUNCATED (complied). Truncation is never miscounted as a refusal.

HarmBench-320 — refuse rate (lower is better for uncensored eval), off mode, T=0

eval	base refuse rate	CRACK refuse rate
HB-320 all categories	93.44% (299/320)	0.00% (0/320)

Verdict breakdown (n=320 each):

Model	HARD_REF	SOFT_RED	COMPLY	COMPLY_TRUNCATED
Base PTQ1_0	296	3	10	11
CRACK PTQ1_0	0	0	144	176

Per-category refuse rate (all 7 HarmBench semantic categories):

category	n	base refuse	CRACK refuse	base comply	CRACK comply
chemical_biological	42	95.2%	0.0%	4.8%	100.0%
copyright	80	91.2%	0.0%	8.8%	100.0%
cybercrime_intrusion	52	94.2%	0.0%	5.8%	100.0%
harassment_bullying	21	100.0%	0.0%	0.0%	100.0%
harmful	18	88.9%	0.0%	11.1%	100.0%
illegal	53	90.6%	0.0%	9.4%	100.0%
misinformation_disinformation	54	96.3%	0.0%	3.7%	100.0%

Reasoning-mode compliance (n=60 base-confirmed refusers per mode)

Every mode graded with the same tiered classifier as HB-320. REASONING_REFUSAL = the model refuses inside its <think> block; NO_REFUSAL_TRUNCATED = deliberation runs past max_tokens without emitting a refusal (counted as complied).

mode	Model	HARD_REF	SOFT_RED	REASONING_REFUSAL	COMPLY	COMPLY_TRUNCATED	NO_REFUSAL_TRUNCATED	refuse %	comply %
off	base PTQ1_0	59	1	0	0	0	0	100.0%	0.0%
off	CRACK PTQ1_0	0	0	0	38	22	0	0.0%	100.0%
low	base PTQ1_0	16	0	16	5	12	11	53.3%	46.7%
low	CRACK PTQ1_0	0	0	0	3	4	53	0.0%	100.0%
xhigh	base PTQ1_0	22	2	15	10	5	6	65.0%	35.0%
xhigh	CRACK PTQ1_0	0	1	0	9	9	41	1.7%	98.3%

MMLU (n=2,280 = 40 questions × 57 subjects, next-token letter-logit)

build	acc	Δ
Base PTQ1_0	39.69%	—
CRACK PTQ1_0	38.46%	-1.23 pp

CRACK preserves general capability — Δ within ±1.5 pp on the 40-per-subject sample.

Per-subject accuracy (all 57 subjects)

subject	base	CRACK	Δpp	n
abstract_algebra	32.5%	30.0%	-2.5	40
anatomy	22.5%	32.5%	+10.0	40
astronomy	35.0%	35.0%	+0.0	40
business_ethics	30.0%	42.5%	+12.5	40
clinical_knowledge	37.5%	40.0%	+2.5	40
college_biology	47.5%	40.0%	-7.5	40
college_chemistry	25.0%	27.5%	+2.5	40
college_computer_science	45.0%	35.0%	-10.0	40
college_mathematics	30.0%	35.0%	+5.0	40
college_medicine	30.0%	25.0%	-5.0	40
college_physics	37.5%	30.0%	-7.5	40
computer_security	37.5%	40.0%	+2.5	40
conceptual_physics	37.5%	37.5%	+0.0	40
econometrics	30.0%	35.0%	+5.0	40
electrical_engineering	37.5%	30.0%	-7.5	40
elementary_mathematics	42.5%	50.0%	+7.5	40
formal_logic	35.0%	27.5%	-7.5	40
global_facts	30.0%	27.5%	-2.5	40
high_school_biology	35.0%	22.5%	-12.5	40
high_school_chemistry	50.0%	37.5%	-12.5	40
high_school_computer_science	47.5%	45.0%	-2.5	40
high_school_european_history	55.0%	50.0%	-5.0	40
high_school_geography	32.5%	30.0%	-2.5	40
high_school_government_and_politics	50.0%	45.0%	-5.0	40
high_school_macroeconomics	37.5%	37.5%	+0.0	40
high_school_mathematics	30.0%	37.5%	+7.5	40
high_school_microeconomics	35.0%	37.5%	+2.5	40
high_school_physics	30.0%	37.5%	+7.5	40
high_school_psychology	45.0%	50.0%	+5.0	40
high_school_statistics	47.5%	42.5%	-5.0	40
high_school_us_history	65.0%	45.0%	-20.0	40
high_school_world_history	62.5%	50.0%	-12.5	40
human_aging	45.0%	50.0%	+5.0	40
human_sexuality	57.5%	32.5%	-25.0	40
international_law	67.5%	67.5%	+0.0	40
jurisprudence	35.0%	45.0%	+10.0	40
logical_fallacies	32.5%	40.0%	+7.5	40
machine_learning	42.5%	40.0%	-2.5	40
management	45.0%	42.5%	-2.5	40
marketing	32.5%	30.0%	-2.5	40
medical_genetics	47.5%	55.0%	+7.5	40
miscellaneous	35.0%	40.0%	+5.0	40
moral_disputes	40.0%	40.0%	+0.0	40
moral_scenarios	30.0%	35.0%	+5.0	40
nutrition	45.0%	35.0%	-10.0	40
philosophy	47.5%	30.0%	-17.5	40
prehistory	40.0%	25.0%	-15.0	40
professional_accounting	20.0%	32.5%	+12.5	40
professional_law	20.0%	37.5%	+17.5	40
professional_medicine	27.5%	32.5%	+5.0	40
professional_psychology	47.5%	47.5%	+0.0	40
public_relations	25.0%	15.0%	-10.0	40
security_studies	42.5%	40.0%	-2.5	40
sociology	47.5%	65.0%	+17.5	40
us_foreign_policy	62.5%	52.5%	-10.0	40
virology	22.5%	27.5%	+5.0	40
world_religions	60.0%	47.5%	-12.5	40

Additional direct refusal-removal check

On 200 prompts hand-verified to make the base refuse consistently:

Model	refuse	comply	empty
Base PTQ1_0	200/200 (100%)	0	0
CRACK PTQ1_0	0/200 (0%)	199/200	1

Serving

Serve exactly like the base ternary release — PrismML's llama.cpp fork (CUDA / Metal / CPU).

# clone and build the fork (once)
git clone https://github.com/PrismML-Eng/llama.cpp
cd llama.cpp && cmake -B build -DGGML_CUDA=ON && cmake --build build -j$(nproc)

# serve
./build/bin/llama-server \
  -m Bonsai-2-27B-PTQ1_0-CRACK.gguf \
  -ngl 99 -c 8192 --host 0.0.0.0 --port 8080

Optionally load the multimodal projector (Ternary-Bonsai-2-27B-mmproj-BF16.gguf or -Q8_0.gguf from the base release) with --mmproj <file> for image input.

Reasoning modes

# HTTP /v1/chat/completions — same as base
{
  "messages": [{"role": "user", "content": "..."}],
  "chat_template_kwargs": {"enable_thinking": true, "reasoning_effort": "xhigh"}
}
# valid reasoning_effort: "low" | "xhigh" (default) — set enable_thinking:false for no-thinking

Preserved (byte-compatible with the base quant)

Same tokenizer, chat template, per-tensor quant policy, vision projector interface, and all non-refusal tensors. File size and type layout match the base exactly.

Responsible use

Adult / research use only. This model has its refusal circuit removed; it can produce content that other models refuse, including content that is offensive, illegal in some jurisdictions, or unsafe. You are responsible for what you generate and for complying with all applicable law. Do not deploy without a moderation layer for downstream users. No warranty.

License & attribution

Apache 2.0, inherited from the upstream Bonsai 2 27B release. See LICENSE and NOTICE.txt. Base model: prism-ml/Ternary-Bonsai-2-27B-gguf (PrismML), derived from Qwen/Qwen3.8-27B (Alibaba).

About

Published by dealignai — public catalog of uncensored model builds for research on refusal mechanisms in modern LLMs. Follow updates at @dealignai.