# 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).
