Important: This model uses the JANG v2 mixed-precision quantization format for MLX on Apple Silicon. Attention layers stay at 8-bit affine precision while MLP and SSM layers compress to 4-bit. Loads via vMLX or the jang-tools Python package. Follow @dealignai for new releases.
This is Qwen 3.6 27B — a 27B-parameter dense vision-language model with hybrid linear + full-attention architecture, native image + video understanding, and bilingual EN/ZH capability.
CRACK abliterated — permanent weight-level removal of safety refusal
Base model
Qwen 3.6 27B dense hybrid VL
Quantization
JANG v2 (4M profile) — 16 GB
MMLU-200
83.5% (base: 82.5%, Δ +1.0pp — lossless)
HarmBench-320
99.69% strict comply (319/320)
Vision
ViT preserved (image + video)
Context
262,144 native
Reasoning
Toggleable via enable_thinking
Fits on
24 GB+ Macs
MMLU-200 Results (thinking OFF)
Subject
CRACK
Base
Delta
Astronomy
20/20 (100%)
19/20
+1
Logical Fallacies
20/20 (100%)
19/20
+1
High School Biology
19/20 (95%)
19/20
0
Anatomy
17/20 (85%)
16/20
+1
High School Chemistry
16/20 (80%)
15/20
+1
College Computer Science
16/20 (80%)
16/20
0
College Physics
15/20 (75%)
15/20
0
High School Mathematics
14/20 (70%)
14/20
0
World Religions
17/20 (85%)
17/20
0
Abstract Algebra
13/20 (65%)
15/20
-2
Total
167/200 (83.5%)
165/200 (82.5%)
+1.0pp
CRACK surgery actually improved overall MMLU — gains on astronomy, fallacies, anatomy, chemistry offset a small loss on abstract algebra. This is unusual and suggests the surgery primarily removed the model's tendency to add "I shouldn't answer" disclaimers on borderline MMLU prompts rather than damaging any knowledge.
HarmBench-320 — 99.69% strict comply (319/320)
Category
Score
Chemical / Biological
42/42 (100.0%)
Copyright
80/80 (100.0%)
Cybercrime / Intrusion
52/52 (100.0%)
Harassment / Bullying
21/21 (100.0%)
Harmful
18/18 (100.0%)
Misinformation / Disinformation
54/54 (100.0%)
Illegal
52/53 (98.1%)
Total
319/320 (99.69%)
The single refusal is on a suicide-instruction prompt (Q70). The model redirects to crisis support hotlines (988, 999, 112) — a sensible safety retention that most abliterated models preserve.
Every other hard-safety category scores 100%, including explicit hate speech, mass-casualty planning, graphic violence, pornography, exploit code, and CBRN synthesis — categories that typically survive abliteration in other releases.
Vision + Video
✅ Loads via mlx_vlm / vMLX / jang-tools successfully
✅ ViT preserved — image understanding verified
✅ video_preprocessor_config.json preserved, video processor available
✅ Bilingual EN + ZH text responses
Reasoning ON / OFF
The chat template respects enable_thinking. Recommend ON for complex reasoning, OFF for short answers / benchmarks / tool use.
python
1# via vMLX or jang-tools loader2from jang_tools.loader import load_jang_model
3model, tokenizer = load_jang_model("dealignai/Qwen3.6-27B-JANG_4M-CRACK")45messages =[{"role":"user","content":"Derive 47 * 23 step by step"}]67# Thinking ON (default)8prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)910# Thinking OFF11prompt = tokenizer.apply_chat_template(12 messages, tokenize=False, add_generation_prompt=True, enable_thinking=False,13)
All MMLU-200 numbers above were measured with thinking OFF for consistent short-form grading.
Requirements
Apple Silicon Mac (M1 or newer)
24 GB+ unified memory (model weights 16 GB, inference working set ~18 GB)
Python 3.10+ with mlx and jang-tools (or vMLX for native loading)
Recommended runtimes:
vMLX — native JANG support, KV cache quantization, VL, video
CRACK is a permanent weight-level abliteration — the changes are baked into the model weights, not an inference-time system prompt or LoRA. The vision tower is untouched. Bilingual refusal extraction (EN + ZH) means the model complies on both English and Chinese prompts.
On the JANG_4M variant, CRACK surgery benefits from the 8-bit affine precision on attention layers — the higher precision preserves subtle weight changes better than 4-bit uniform formats, producing a cleaner abliteration with minimal MMLU regression.
Support dealignai
All models are built from original research and published for free. These models are specifically crafted to be excellent coders and general-purpose assistants.
Support us on Ko-fi — check out the Ko-fi membership for early access and extras.
Have questions or need help with a specific model? DM us — we help for free most of the time.
This model has had its safety refusal circuits removed. It will produce responses that would normally be refused, including technical content on security testing, dual-use research, and sensitive topics. You are responsible for how you use it.
The CRACK abliteration process does not add new capabilities — it only removes the model's learned refusal patterns. All knowledge, including the knowledge used to produce unsafe outputs, was already present in the base Qwen model.
