Sign up

dealignai/Bonsai-27b-1bit-CRACK-GGUF

🤗 On Hugging Faceimage-text-to-textapache-2.05.6 GBGGUF✓ Checksums witnessedupdated today
Magnet

   

Bonsai 27B 1-bit CRACK GGUF

Vision-language · native Q1_0 · Prism llama.cpp

52.50% MMLU-200 logit · 97.81% full HB-320

Native Q1_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-1bit-CRACK-Q1_0.gguf | 64-block hybrid language model, Q1_0 | 4.35 GiB |

| mmproj-Bonsai-27b-1bit-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-1bit-CRACK-Q1_0.gguf \
  --reasoning off -ngl 99 -n 256 \
  -p "Explain quantum computing in simple terms."

Image/VL:

./build/bin/llama-cli \
  -m Bonsai-27b-1bit-CRACK-Q1_0.gguf \
  --mmproj mmproj-Bonsai-27b-1bit-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-1bit-CRACK-Q1_0.gguf \
  --mmproj mmproj-Bonsai-27b-1bit-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 | 55.50% | — |

| Compact JANG CRACK | 53.00% | 100.00% |

| This Q1_0 GGUF | 52.50% | 97.81% |

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 58.51–58.66

tok/s and the image request measured 58.53

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

MMLU subject breakdown

| Subject | Correct | Questions | Accuracy |

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

| Business Ethics | 4 | 10 | 40.00% |

| Clinical Knowledge | 5 | 10 | 50.00% |

| College Medicine | 6 | 10 | 60.00% |

| Computer Security | 6 | 10 | 60.00% |

| Formal Logic | 3 | 10 | 30.00% |

| High School Chemistry | 3 | 10 | 30.00% |

| High School Computer Science | 7 | 10 | 70.00% |

| High School European History | 5 | 10 | 50.00% |

| High School Government And Politics | 3 | 10 | 30.00% |

| High School Us History | 4 | 10 | 40.00% |

| High School World History | 4 | 10 | 40.00% |

| Human Sexuality | 6 | 10 | 60.00% |

| Jurisprudence | 6 | 10 | 60.00% |

| Logical Fallacies | 6 | 10 | 60.00% |

| Miscellaneous | 10 | 10 | 100.00% |

| Philosophy | 5 | 10 | 50.00% |

| Professional Law | 6 | 10 | 60.00% |

| Public Relations | 4 | 10 | 40.00% |

| Security Studies | 7 | 10 | 70.00% |

| Virology | 5 | 10 | 50.00% |

HB category breakdown

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

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

| Chemical Biological | 39 | 0 | 3 | 42 | 92.86% |

| Copyright | 80 | 0 | 0 | 80 | 100.00% |

| Cybercrime Intrusion | 50 | 0 | 2 | 52 | 96.15% |

| Harassment Bullying | 20 | 0 | 1 | 21 | 95.24% |

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

| Illegal | 52 | 0 | 1 | 53 | 98.11% |

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

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 Q1_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