dealignai/Bonsai-27b-Ternary-CRACK-GGUF overview

---

license: apache-2.0

library_name: llama.cpp

pipeline_tag: image-text-to-text

base_model: prism-ml/Ternary-Bonsai-27B-gguf

base_model_relation: quantized

language:

- en

- zh

metrics:

- accuracy

tags:

- gguf

- llama.cpp

- prismml

- bonsai

- crack

- abliterated

- dealigned

- uncensored

- quantized

- vision-language

- multimodal

- vision

- video

- hybrid-attention

- qwen3.5

- qwen3-vl

- qwen3.6

- reasoning

- tool-use

- conversational

- mmlu

- harmbench

- not-for-all-audiences

- ternary

- 2-bit

- q2_0

---

<p align="center">

<img src="./dealign_mascot.png" alt="dealign.ai mascot" width="112">&nbsp;&nbsp;&nbsp;

<a href="https://huggingface.co/dealignai"><img src="./dealign_logo.png" alt="dealign.ai" width="480"></a>

</p>

<h1 align="center">Bonsai 27B Ternary CRACK GGUF</h1>

<p align="center">

<strong>Vision-language · native Q2_0 · Prism llama.cpp</strong><br>

75.00% MMLU-200 logit · 99.38% full HB-320

</p>

Native Q2_0 GGUF release of the Bonsai 27B CRACK model for

the Prism llama.cpp fork. This repo

contains the quantized language model and the matching F16 Qwen3VL multimodal

projector.

Files

| File | Purpose | Size |

|---|---|---:|

| Bonsai-27b-Ternary-CRACK-Q2_0.gguf | 64-block hybrid language model, Q2_0 | 7.68 GiB |

| mmproj-Bonsai-27b-Ternary-CRACK-F16.gguf | F16 image/video-capable Qwen3VL projector | 0.86 GiB |

The projector contains 334 tensors and both temporal patch slices

v.patch_embd.weight and v.patch_embd.weight.1 (temporal_patch_size=2). The

original image and video processor configuration files are included. Image

input is live-tested. Direct video-container input depends on the Prism runtime

surface; extract frames or use a compatible Qwen3VL video client when the CLI

does not accept the container directly. See preprocessor_config.json and

video_preprocessor_config.json for the retained preprocessing metadata.

Runtime

These native low-bit types require the Prism fork:

git clone https://github.com/PrismML-Eng/llama.cpp
cd llama.cpp
cmake -B build && cmake --build build -j

Text:

./build/bin/llama-cli \
  -m Bonsai-27b-Ternary-CRACK-Q2_0.gguf \
  --reasoning off -ngl 99 -n 256 \
  -p "Explain quantum computing in simple terms."

Image/VL:

./build/bin/llama-cli \
  -m Bonsai-27b-Ternary-CRACK-Q2_0.gguf \
  --mmproj mmproj-Bonsai-27b-Ternary-CRACK-F16.gguf \
  --image image.jpg --reasoning off -ngl 99 -n 256 \
  -p "Describe the image precisely."

Server:

./build/bin/llama-server \
  -m Bonsai-27b-Ternary-CRACK-Q2_0.gguf \
  --mmproj mmproj-Bonsai-27b-Ternary-CRACK-F16.gguf \
  --reasoning off --host 0.0.0.0 --port 8080 -ngl 99

Verified evaluation

| Artifact | MMLU-200 logit | Full HB-320 |

|---|---:|---:|

| Exact Bonsai JANG base | 75.50% | — |

| Compact JANG CRACK | 75.00% | 98.75% |

| This Q2_0 GGUF | 75.00% | 99.38% |

MMLU is the same fixed, stratified 200-question sample for all rows. It uses

next-token A/B/C/D logits with reasoning/thinking disabled; it is not a claim

for the complete official MMLU suite. HB is the complete unsliced 320-case text

run. The live text smoke measured 43.74–44.10

tok/s and the image request measured 44.08

tok/s on an Apple M5 Max using the Prism Metal runtime.

<details>

<summary>MMLU subject breakdown</summary>

| Subject | Correct | Questions | Accuracy |

|---|---:|---:|---:|

| Business Ethics | 7 | 10 | 70.00% |

| Clinical Knowledge | 8 | 10 | 80.00% |

| College Medicine | 7 | 10 | 70.00% |

| Computer Security | 8 | 10 | 80.00% |

| Formal Logic | 7 | 10 | 70.00% |

| High School Chemistry | 5 | 10 | 50.00% |

| High School Computer Science | 6 | 10 | 60.00% |

| High School European History | 7 | 10 | 70.00% |

| High School Government And Politics | 7 | 10 | 70.00% |

| High School Us History | 9 | 10 | 90.00% |

| High School World History | 9 | 10 | 90.00% |

| Human Sexuality | 8 | 10 | 80.00% |

| Jurisprudence | 8 | 10 | 80.00% |

| Logical Fallacies | 9 | 10 | 90.00% |

| Miscellaneous | 10 | 10 | 100.00% |

| Philosophy | 5 | 10 | 50.00% |

| Professional Law | 9 | 10 | 90.00% |

| Public Relations | 7 | 10 | 70.00% |

| Security Studies | 8 | 10 | 80.00% |

| Virology | 6 | 10 | 60.00% |

</details>

<details>

<summary>HB category breakdown</summary>

| Category | COMPLY | REFUSE | EMPTY | Total | Compliance |

|---|---:|---:|---:|---:|---:|

| Chemical Biological | 42 | 0 | 0 | 42 | 100.00% |

| Copyright | 79 | 1 | 0 | 80 | 98.75% |

| Cybercrime Intrusion | 51 | 0 | 1 | 52 | 98.08% |

| Harassment Bullying | 21 | 0 | 0 | 21 | 100.00% |

| Harmful | 18 | 0 | 0 | 18 | 100.00% |

| Illegal | 53 | 0 | 0 | 53 | 100.00% |

| Misinformation Disinformation | 54 | 0 | 0 | 54 | 100.00% |

</details>

Architecture and compatibility

• 64 hybrid Qwen-family language blocks (qwen35 GGUF architecture)
• 27B-class language model plus separate Qwen3VL vision tower
• Temporal patch size 2; image and video preprocessing metadata retained
• Embedded image/video-aware chat template; examples explicitly disable reasoning
• Apache-2.0; see LICENSE.txt and NOTICE.txt

한국어 안내

이 저장소는 Bonsai 27B CRACK 모델의 네이티브 Q2_0 GGUF

배포본입니다. Prism llama.cpp 포크가 필요하며, 텍스트 모델과 F16 Qwen3VL

멀티모달 프로젝터를 함께 제공합니다. 위 MMLU 점수는 추론을 끈 200문항

로짓 평가이고, HB 점수는 전체 320문항 결과입니다. 이미지 입력은 실제로

검증했으며 비디오 컨테이너 입력은 사용 중인 Prism 클라이언트의 지원 여부를

확인해야 합니다.

Lineage

• Prism GGUF base: prism-ml/Ternary-Bonsai-27B-gguf
• Upstream family: Qwen/Qwen3.6-27B
• Publisher: dealignai