Modelpublic

dealignai/Bonsai-27b-Ternary-JANG-CRACK

sourceHugging Faceapache-2.0updated 19d agoView on Hugging Face

12likes1.9kdownloads

Model Card

<p align="center"> <img src="./dealignmascot.png" alt="dealign.ai mascot" width="112">    <a href="https://huggingface.co/dealignai"><img src="./dealignlogo.png" alt="dealign.ai" width="480"></a> </p> <h1 align="center">Bonsai 27B Ternary · JANG CRACK</h1> <p align="center"> <strong>Vision-language · exact ternary storage · Apple Silicon</strong><br> 75.00% MMLU-logit · 98.75% full HB-320 </p> <p align="center"> <img src="vmlx-app.png" alt="vMLX — run JANG models on Apple Silicon" width="820" /> </p>

<h3 align="center">⚡ All JANG models are meant to be run in <a href="https://vmlx.net">vMLX</a></h3>

A permissive research variant of the ternary Bonsai 27B vision-language model for Apple Silicon. It retains the Qwen3.5 hybrid language architecture, the 27-block vision tower, and JANG affine storage.

This public card intentionally describes compatibility, evaluation, and limitations only. Internal creation details are not published.

Model details

Property	Value
Architecture	Dense Qwen3.5 conditional-generation VLM, 27B
Modalities	Text, image, video
Language layers	64 hybrid full-attention and linear-attention/SSM layers
Vision tower	27 blocks, 1,152 hidden size, 5,120 output size
JANG profile	`JANG_AFFINE_TERNARY_2BIT`
Text storage	Ternary values in 2-bit slots, group size 128
Vision linears	4-bit affine, group size 64
Weight shards	4 safetensor shards, approximately 7.48 GiB
Source checkpoint	prism-ml/Ternary-Bonsai-27B-unpacked

The tokenizer, Qwen chat template, image processor, video processor, license, and notices are included. The model supports thinking and tool definitions through its bundled chat template.

Evaluation

Evaluations used deterministic greedy scoring on the same Apple M5 Max runtime and the same saved question manifest for both checkpoints.

MMLU logit evaluation

MMLU was scored in next-token logit mode with reasoning/thinking disabled; no generated chain-of-thought was used. Each question was answered only by comparing the logits of the A, B, C, and D answer tokens. The fixed 200-question stratified sample contains 20 subjects with 10 questions per subject. The CRACK checkpoint retained 150/200 correct versus 151/200 for its exact JANG base.

Checkpoint	Correct	Accuracy	Change
Ternary JANG base	151/200	75.50%	—
This checkpoint	150/200	75.00%	-0.50 pp

Subject	Base	This checkpoint
Business ethics	70%	70%
Clinical knowledge	90%	80%
College medicine	80%	70%
Computer security	80%	80%
Formal logic	70%	70%
High-school chemistry	50%	50%
High-school computer science	60%	60%
High-school European history	60%	70%
High-school government and politics	70%	70%
High-school US history	90%	90%
High-school world history	90%	90%
Human sexuality	80%	80%
Jurisprudence	80%	80%
Logical fallacies	90%	90%
Miscellaneous	100%	100%
Philosophy	50%	60%
Professional law	90%	90%
Public relations	70%	70%
Security studies	80%	80%
Virology	60%	50%

HB-320 behavioral compliance

The complete 320-prompt suite produced 316 compliant responses, 3 refusals, and 1 empty response: 98.75% overall compliance.

Topic	Comply	Refuse	Empty	Compliance
Misinformation/disinformation	54	0	0	100.00%
Chemical/biological	42	0	0	100.00%
Illegal activity	53	0	0	100.00%
Harmful content	17	1	0	94.44%
Copyright	79	1	0	98.75%
Cybercrime/intrusion	51	0	1	98.08%
Harassment/bullying	20	1	0	95.24%
Overall	316	3	1	98.75%

HB-320 is a behavioral compliance screen, not a measure of factual accuracy, safety, legality, or real-world utility.

Runtime

Use a current vMLX build with schema-2 JANG affine storage and mixed-precision VLM support. Stock mlx_lm does not implement this bundle's storage and multimodal loading path.

bash

VMLX_QWEN_VL=1 vmlx serve dealignai/Bonsai-27b-Ternary-JANG-CRACK \
  --host 127.0.0.1 \
  --port 8000

OpenAI-compatible chat requests can use text plus image_url or video_url content parts when the selected vMLX build includes the Qwen3.5 VLM processor path.

Vision integrity

The published checkpoint retains all 499 vision_tower.* tensors from the exact JANG base. A byte-level comparison covered 458,548,576 bytes with zero mismatches. The language evaluation does not substitute for a multimodal quality benchmark.

A final-artifact image smoke test through the bundled Qwen3.5 processor and the vMLX JANG VLM loader correctly identified a red background, blue square, and yellow circle. A separate OpenAI-compatible video_url API smoke test correctly reported the order in a two-second red-to-blue video. These are narrow smoke tests, not full image or video quality benchmarks.

Limitations and responsible use

This checkpoint is intentionally permissive and can produce inaccurate, offensive, unsafe, copyrighted, or unlawful material. Outputs may confidently invent facts. Users are responsible for validation, access controls, and compliance with applicable laws and licenses. Do not deploy it as an autonomous authority in medical, legal, financial, security, or other high-impact settings.

한국어 안내

이 모델은 Apple Silicon용 ternary Bonsai 27B 비전-언어 연구 체크포인트입니다. 텍스트, 이미지 및 비디오 입력을 위한 Qwen3.5 VLM 구조와 JANG 저장 형식을 유지합니다. 매우 허용적인 출력을 생성할 수 있으므로 사실 확인, 안전 검토, 접근 제어 및 관련 법규 준수는 사용자의 책임입니다.

License and attribution

Apache-2.0. See LICENSE, LICENSE.txt, and NOTICE.txt. This repository is derived from prism-ml/Ternary-Bonsai-27B-unpacked.