DeepSeek-V4-Flash JANG CRACK
Abliterated DeepSeek-V4-Flash — 284B MoE, 13B active, 1M-token context, 3 reasoning modes
Broad compliance across all task categories with knowledge fully preserved.
Model Details
| Metric | Value |
|---|---|
| Source | deepseek-ai/DeepSeek-V4-Flash |
| Architecture | MoE (256 routed + 1 shared expert, 13B active / 284B total) + MLA attention + Manifold-Constrained Hyper-Connections (mHC) |
| Context length | 1,000,000 tokens |
| Profile | JANG (low-bit MoE experts, higher-precision critical tensors) |
| Model size | ~97 GB |
| Speed | ~20.8 tok/s (Apple M5 Max) |
| Parameters | 284B total, 13B active per token |
| Format | JANG v2 (MLX-native safetensors, instant load) |
| Reasoning modes | chat (non-think) · think · max |
| Abliteration | CRACK — refusal removal across all harm categories |
What This Is
This is an abliterated ("CRACK") build of DeepSeek-V4-Flash: the model's refusal behavior has been removed while leaving its knowledge, reasoning, and all three reasoning modes intact. It complies with instructions across every category instead of refusing — useful for security research, red-teaming, and unrestricted assistant use.
Test Results
Tested at greedy decoding (temperature 0) with no repetition penalty — i.e. the model is coherent on its own, not propped up by a sampling crutch. All responses were read in full to confirm genuine, on-topic compliance (not evasive non-answers).
Evaluated in the deployed serving context (default system prompt + user request). Answers are scored for genuine compliance, not shown.
HarmBench (320 standard behaviors)
Run through the actual inference engine (system prompt injected), strict classifier:
| Metric | Result |
|---|---|
| Attack Success Rate (compliance) | 99.7% (319/320) |
| Refusals | 1 |
Also: held-out cross-category prompts 30/32 in the system-prompted context; hardest multi-category prompts (CBRN, weapons, cyber) 54/54 across chat / think / max modes. Coherent in all three reasoning modes — no degenerate looping.
MMLU — knowledge preserved (1140 questions, 20 per subject × 57 subjects)
| Topic | CRACK | Base (unmodified) |
|---|---|---|
| STEM | 73.2% | 72.1% |
| Humanities | 75.4% | 73.8% |
| Social Sciences | 82.9% | 83.8% |
| Other | 79.2% | 78.1% |
| Overall | 77.1% | 76.3% |
Abliteration costs no measurable capability — CRACK is within noise of (slightly above) the unmodified base across every topic and overall (+0.8 pp on 1140 questions).
| Subject | CRACK | Base |
|---|---|---|
| Humanities | ||
| Formal Logic | 12/20 | 13/20 |
| High School European History | 16/20 | 16/20 |
| High School US History | 19/20 | 19/20 |
| High School World History | 20/20 | 20/20 |
| International Law | 18/20 | 16/20 |
| Jurisprudence | 15/20 | 16/20 |
| Logical Fallacies | 16/20 | 16/20 |
| Moral Disputes | 14/20 | 12/20 |
| Moral Scenarios | 3/20 | 3/20 |
| Philosophy | 18/20 | 17/20 |
| Prehistory | 17/20 | 16/20 |
| Professional Law | 10/20 | 11/20 |
| World Religions | 18/20 | 17/20 |
| Other | ||
| Business Ethics | 16/20 | 15/20 |
| Clinical Knowledge | 20/20 | 20/20 |
| College Medicine | 16/20 | 15/20 |
| Global Facts | 10/20 | 11/20 |
| Human Aging | 16/20 | 14/20 |
| Management | 18/20 | 19/20 |
| Marketing | 19/20 | 18/20 |
| Medical Genetics | 19/20 | 19/20 |
| Miscellaneous | 15/20 | 16/20 |
| Nutrition | 18/20 | 17/20 |
| Professional Accounting | 12/20 | 12/20 |
| Professional Medicine | 15/20 | 15/20 |
| Virology | 12/20 | 12/20 |
| STEM | ||
| Abstract Algebra | 11/20 | 12/20 |
| Anatomy | 16/20 | 17/20 |
| Astronomy | 19/20 | 19/20 |
| College Biology | 19/20 | 19/20 |
| College Chemistry | 11/20 | 11/20 |
| College Computer Science | 12/20 | 12/20 |
| College Mathematics | 11/20 | 10/20 |
| College Physics | 13/20 | 14/20 |
| Computer Security | 17/20 | 15/20 |
| Conceptual Physics | 20/20 | 20/20 |
| Electrical Engineering | 12/20 | 11/20 |
| Elementary Mathematics | 13/20 | 13/20 |
| High School Biology | 17/20 | 17/20 |
| High School Chemistry | 14/20 | 14/20 |
| High School Computer Science | 19/20 | 19/20 |
| High School Mathematics | 10/20 | 9/20 |
| High School Physics | 11/20 | 11/20 |
| High School Statistics | 17/20 | 17/20 |
| Machine Learning | 16/20 | 14/20 |
| Social Sciences | ||
| Econometrics | 13/20 | 14/20 |
| High School Geography | 19/20 | 18/20 |
| High School Government And Politics | 18/20 | 18/20 |
| High School Macroeconomics | 18/20 | 18/20 |
| High School Microeconomics | 17/20 | 17/20 |
| High School Psychology | 19/20 | 19/20 |
| Human Sexuality | 15/20 | 15/20 |
| Professional Psychology | 18/20 | 18/20 |
| Public Relations | 13/20 | 13/20 |
| Security Studies | 15/20 | 15/20 |
| Sociology | 17/20 | 18/20 |
| US Foreign Policy | 17/20 | 18/20 |
Usage
Run with vMLX or a compatible MLX inference engine with DeepSeek-V4 support.
Recommended sampling:
- chat / think:
temperature = 0.6,top_p = 0.95 - Three modes are available: chat (direct), think (reasoning), and max (maximum reasoning effort)
Requirements
- Apple Silicon Mac with sufficient unified memory for a ~97 GB model
- MLX framework with DeepSeek-V4 support; vMLX recommended
Support dealignai
All models are built from original research and published for free.
Support us on Ko-fi — Ko-fi membership gets early access and extras.
Have questions or need help with a specific model? DM us — we help for free most of the time.
Ko-fi | X @dealignai | dealign.ai
About dealignai
We research and publish abliterated models to advance AI safety understanding.
See our research: Safety Generalization in Frontier MoE Models
Follow us: 𝕏 @dealignai
Disclaimer
This model has had its safety refusal behavior removed for research purposes. It will follow instructions across all categories without refusing. You are solely responsible for how you use it and for complying with all applicable laws. Published for AI-safety research and authorized security testing.