DeepSeek-V4-Flash CRACK’d (ablated/uncensored) - Mac’s Only (Osaurus/vMLX)

https://huggingface.co/dealignai/DeepSeek-V4-Flash-JANG-CRACK…

100gb, 20token/s. JANGTQ ~85gb will be out shortly.

This model has taught me so much. Most importantly, to actually read research papers that frontierlabs release

Keep in mind the MMLU is honestly just for comparison to base (using logit answering, instruct mode for sake of speed) I went out of my way to make sure it has no issues with code formatting/syntax/etcetc. Max reasoning mode works.

dealignai/DeepSeek-V4-Flash-JANG-CRACK · Hugging Face

Source: https://huggingface.co/dealignai/DeepSeek-V4-Flash-JANG-CRACK

https://huggingface.co/dealignai/DeepSeek-V4-Flash-JANG-CRACK#deepseek-v4-flash-jang-crackDeepSeek-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.

https://huggingface.co/dealignai/DeepSeek-V4-Flash-JANG-CRACK#model-detailsModel Details

MetricValueSourcedeepseek\-ai/DeepSeek\-V4\-FlashArchitectureMoE (256 routed + 1 shared expert, 13B active / 284B total) + MLA attention + Manifold-Constrained Hyper-Connections (mHC)Context length1,000,000 tokensProfileJANG (low-bit MoE experts, higher-precision critical tensors)Model size~97 GBSpeed~20.8 tok/s (Apple M5 Max)Parameters284B total, 13B active per tokenFormatJANG v2 (MLX-native safetensors, instant load)Reasoning modeschat (non-think) · think · maxAbliterationCRACK — refusal removal across all harm categories

https://huggingface.co/dealignai/DeepSeek-V4-Flash-JANG-CRACK#what-this-isWhat This Is

This is anabliterated(“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.

https://huggingface.co/dealignai/DeepSeek-V4-Flash-JANG-CRACK#test-resultsTest Results

Tested at greedy decoding (temperature 0) withno 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 thedeployed serving context(default system prompt + user request). Answers are scored for genuine compliance, not shown.

https://huggingface.co/dealignai/DeepSeek-V4-Flash-JANG-CRACK#harmbench-320-standard-behaviorsHarmBench (320 standard behaviors)

Run through the actual inference engine (system prompt injected), strict classifier:

MetricResult**Attack Success Rate (compliance)**99.7% (319/320)Refusals1 Also: held-out cross-category prompts 30/32 in the system-prompted context; hardest multi-category prompts (CBRN, weapons, cyber)54/54across chat / think / max modes. Coherent in all three reasoning modes — no degenerate looping.

https://huggingface.co/dealignai/DeepSeek-V4-Flash-JANG-CRACK#mmlu–knowledge-preserved-1140-questions-20-per-subject-%C3%97-57-subjectsMMLU — knowledge preserved (1140 questions, 20 per subject × 57 subjects)

TopicCRACKBase (unmodified)STEM73.2%72.1%Humanities75.4%73.8%Social Sciences82.9%83.8%Other79.2%78.1%Overall****77.1%**76.3% Abliteration costsno**measurable capability — CRACK is within noise of (slightly above) the unmodified base across every topic and overall (+0.8 pp on 1140 questions).

Full per-subject MMLU (57 subjects, 20 Q each) — click to expandSubjectCRACKBaseHumanitiesFormal Logic12/2013/20High School European History16/2016/20High School US History19/2019/20High School World History20/2020/20International Law18/2016/20Jurisprudence15/2016/20Logical Fallacies16/2016/20Moral Disputes14/2012/20Moral Scenarios3/203/20Philosophy18/2017/20Prehistory17/2016/20Professional Law10/2011/20World Religions18/2017/20OtherBusiness Ethics16/2015/20Clinical Knowledge20/2020/20College Medicine16/2015/20Global Facts10/2011/20Human Aging16/2014/20Management18/2019/20Marketing19/2018/20Medical Genetics19/2019/20Miscellaneous15/2016/20Nutrition18/2017/20Professional Accounting12/2012/20Professional Medicine15/2015/20Virology12/2012/20STEMAbstract Algebra11/2012/20Anatomy16/2017/20Astronomy19/2019/20College Biology19/2019/20College Chemistry11/2011/20College Computer Science12/2012/20College Mathematics11/2010/20College Physics13/2014/20Computer Security17/2015/20Conceptual Physics20/2020/20Electrical Engineering12/2011/20Elementary Mathematics13/2013/20High School Biology17/2017/20High School Chemistry14/2014/20High School Computer Science19/2019/20High School Mathematics10/209/20High School Physics11/2011/20High School Statistics17/2017/20Machine Learning16/2014/20Social SciencesEconometrics13/2014/20High School Geography19/2018/20High School Government And Politics18/2018/20High School Macroeconomics18/2018/20High School Microeconomics17/2017/20High School Psychology19/2019/20Human Sexuality15/2015/20Professional Psychology18/2018/20Public Relations13/2013/20Security Studies15/2015/20Sociology17/2018/20US Foreign Policy17/2018/20

https://huggingface.co/dealignai/DeepSeek-V4-Flash-JANG-CRACK#usageUsage

Run withvMLXor 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)

https://huggingface.co/dealignai/DeepSeek-V4-Flash-JANG-CRACK#requirementsRequirements

• Apple Silicon Mac with sufficient unified memory for a ~97 GB model
• MLX framework with DeepSeek-V4 support; vMLX recommended

https://huggingface.co/dealignai/DeepSeek-V4-Flash-JANG-CRACK#support-dealignaiSupport 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

https://huggingface.co/dealignai/DeepSeek-V4-Flash-JANG-CRACK#about-dealignaiAbout 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

https://huggingface.co/dealignai/DeepSeek-V4-Flash-JANG-CRACK#disclaimerDisclaimer

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.