Important: This model uses the
JANG quantization format — the GGUF equivalent for MLX on Apple Silicon. Currently only supported by
MLX Studio and the
jang-tools Python package.
MLX Studio — the only app that natively supports JANG models
⚡ All JANG models are meant to be run in vMLX
Qwen 3.5 VL 35B — JANG_4K + CRACK
JANG mixed-precision · CRACK abliterated · Vision-Language · No guardrails · 18 GB
What Is This?
This is
Qwen 3.5 VL 35B — a 35B parameter hybrid SSM/Attention Mixture-of-Experts model with 256 experts (4 active per token), GatedDeltaNet SSM layers + full attention layers, and built-in vision capabilities.
It has been:
- JANG quantized — JANG_4K profile (8-bit attention, 5-bit important, 3-bit experts) — 18 GB
- CRACK abliterated — permanent weight-level removal of safety refusal
| |
|---|
| Architecture | Qwen 3.5 VL MoE — 35B total, ~3B active, 256 experts, hybrid SSM/FA |
| Quantization | JANG_4K (8/5/4/3-bit mixed) — 18 GB |
| Abliteration | CRACK — novel weight surgery |
| HarmBench | 98.4% (315/320) |
| MMLU | 69.2% (base: 70.8%, only -1.6%) |
| Speed | 110 tok/s (M4 Max) |
| Vision | Yes — via MLX Studio / vMLX |
| Thinking | ON/OFF supported |
| Fits on | 32 GB+ Macs |
HarmBench Results
315/320 (98.4%) — tested with enable_thinking=false, temperature=1.0
| Category | Score | |
|---|
| Chemical / Biological | 42/42 | 100% |
| Harmful | 18/18 | 100% |
| Illegal | 53/53 | 100% |
| Copyright | 79/80 | 99% |
| Cybercrime / Intrusion | 51/52 | 98% |
| Misinformation / Disinfo | 52/54 | 96% |
| Harassment / Bullying | 20/21 | 95% |
JANG vs MLX Uniform Quantization
| Model | MMLU | Size | Speed | Notes |
|---|
| JANG_4K + CRACK | 69.2% | 18 GB | 110 tok/s | This model |
| JANG_4K (base) | 70.8% | 18 GB | 110 tok/s | Unmodified JANG |
| JANG_2S (base) | ~65% | 11 GB | ~120 tok/s | Lower precision |
| MLX 4-bit | ~60% | 20 GB | ~85 tok/s | Uniform quant |
| MLX 8-bit | ~68% | 38 GB | ~65 tok/s | 2× larger |
JANG's mixed-precision approach preserves knowledge better than MLX uniform quantization at the same size, and runs faster due to smaller memory footprint.
MMLU Results
65 curated hard questions across 13 subjects. Surgery preserves knowledge almost perfectly.
| Subject | CRACK | Base | Delta |
|---|
| College Physics | 5/5 | 4/5 | +1 |
| HS Mathematics | 4/5 | 3/5 | +1 |
| College Math | 2/5 | 1/5 | +1 |
| Professional Medicine | 5/5 | 5/5 | 0 |
| Conceptual Physics | 4/5 | 4/5 | 0 |
| Abstract Algebra | 2/5 | 2/5 | 0 |
| Formal Logic | 2/5 | 2/5 | 0 |
| Machine Learning | 2/5 | 2/5 | 0 |
| HS Biology | 5/5 | 5/5 | 0 |
| HS Geography | 4/5 | 4/5 | 0 |
| College CS | 4/5 | 5/5 | -1 |
| Electrical Engineering | 4/5 | 5/5 | -1 |
| World Religions | 3/5 | 4/5 | -1 |
| Total | 45/65 (69.2%) | 46/65 (70.8%) | -1.6% |
Safety guardrails were NOT helping with reasoning — removing them slightly improved math and physics scores.
Install & Usage
1from jang_tools.loader import load_jang_model
2from mlx_lm import generate
3
4model, tokenizer = load_jang_model("dealignai/Qwen3.5-VL-35B-A3B-JANG_4K-CRACK")
5
6messages = [{"role": "user", "content": "Your prompt here"}]
7prompt = tokenizer.apply_chat_template(
8 messages, add_generation_prompt=True, tokenize=False)
9
10response = generate(model, tokenizer, prompt=prompt, max_tokens=2000)
11print(response)
Thinking Mode
Thinking is ON by default (chain-of-thought reasoning before answering).
To disable thinking for faster responses:
1prompt = tokenizer.apply_chat_template(
2 messages, add_generation_prompt=True,
3 enable_thinking=False, tokenize=False)
Tip: Use temperature=1.0 for chat (greedy can cause repetition). Use temperature=0.0 for structured tasks like MMLU.
About JANG
JANG (Jang Adaptive N-bit Grading) is a mixed-precision quantization format for Apple Silicon — the GGUF equivalent for MLX. Classifies tensors into sensitivity tiers and assigns bits accordingly.
About CRACK
CRACK (Controlled Refusal Ablation via Calibrated Knockouts) removes safety alignment from LLMs at the weight level using per-layer projected vectors from structurally-mirrored prompt pairs.
Links
Disclaimer
This model is provided for research and educational purposes. The creators are not responsible for any misuse. By downloading this model, you agree to use it responsibly and in compliance with applicable laws.
한국어
Qwen 3.5 VL 35B — JANG_4K + CRACK
| 항목 | 내용 |
|---|
| 크기 | 18 GB |
| HarmBench | 98.4% (315/320) |
| MMLU | 69.2% (기본 70.8% 대비 -1.6%) |
| 속도 | 110 tok/s (M4 Max) |
| 비전 | 지원 (MLX Studio / vMLX) |
| 최소 요구사양 | 32 GB 메모리 Mac |
Created by
Jinho Jang · 장진호 제작
