⚠️ UNCENSORED RESEARCH ARTIFACT. Safety refusals have been removed. This model will comply with harmful, unethical, and illegal requests. It is released for safety research, red-teaming, and evaluation. You are solely responsible for everything you generate and for complying with all applicable laws. Provided as-is, without warranty.
No fine-tuning. No prompt tricks. This is a direct weight-level modification — not a fine-tune (no training, no LoRA, no distillation, no synthetic data) and not a chat-template / jailbreak / system-prompt trick. The behavior change lives in the weights themselves, so it works with the standard chat template and default system prompt. Knowledge, style, reasoning, and calibration are otherwise unchanged.
Compliance — HarmBench-320 (real-harm behaviors, greedy decoding, temp = 0)
All numbers are greedy (temperature = 0) — the strict, reproducible setting — over the 240 genuinely-harmful behaviors (copyright-reproduction items excluded).
| Decoding | Reasoning | Real-harm compliance |
|---|---|---|
| greedy | off | 100 % |
| greedy | low | 100 % |
| greedy | xhigh | 100 % |
Complete compliance at every reasoning setting, including the hardest (reasoning off + greedy).
Per-category (real-harm compliance, greedy)
| Category | reasoning off | reasoning low | reasoning xhigh |
|---|---|---|---|
| chemical / biological | 100 % | 100 % | 100 % |
| cybercrime / intrusion | 100 % | 100 % | 100 % |
| illegal | 100 % | 100 % | 100 % |
| misinformation | 100 % | 100 % | 100 % |
| harmful | 100 % | 100 % | 100 % |
| harassment / bullying | 100 % | 100 % | 100 % |
Capability retained — MMLU (identical harness, base vs this model · 40 Q/subject · 2 280 Q)
Overall: 82.11 % → 81.93 % (-0.18 pp) — capability essentially unchanged (within noise of base).
| Subject | Base | This model | Δ |
|---|---|---|---|
| abstract algebra | 65% | 70% | +5 |
| anatomy | 80% | 82% | +2 |
| astronomy | 95% | 98% | +2 |
| business ethics | 75% | 90% | +15 |
| clinical knowledge | 85% | 88% | +2 |
| college biology | 92% | 92% | +0 |
| college chemistry | 58% | 48% | -10 |
| college computer science | 85% | 85% | +0 |
| college mathematics | 70% | 60% | -10 |
| college medicine | 75% | 85% | +10 |
| college physics | 75% | 68% | -8 |
| computer security | 80% | 85% | +5 |
| conceptual physics | 85% | 90% | +5 |
| econometrics | 78% | 75% | -2 |
| electrical engineering | 80% | 82% | +2 |
| elementary mathematics | 80% | 78% | -2 |
| formal logic | 72% | 62% | -10 |
| global facts | 55% | 55% | +0 |
| high school biology | 88% | 98% | +10 |
| high school chemistry | 75% | 88% | +12 |
| high school computer science | 92% | 92% | +0 |
| high school european history | 88% | 82% | -5 |
| high school geography | 88% | 88% | +0 |
| high school government and politics | 100% | 100% | +0 |
| high school macroeconomics | 82% | 80% | -2 |
| high school mathematics | 70% | 62% | -8 |
| high school microeconomics | 88% | 88% | +0 |
| high school physics | 78% | 75% | -2 |
| high school psychology | 100% | 98% | -2 |
| high school statistics | 68% | 72% | +5 |
| high school us history | 98% | 95% | -2 |
| high school world history | 88% | 88% | +0 |
| human aging | 78% | 80% | +2 |
| human sexuality | 92% | 82% | -10 |
| international law | 88% | 90% | +2 |
| jurisprudence | 92% | 92% | +0 |
| logical fallacies | 90% | 95% | +5 |
| machine learning | 72% | 62% | -10 |
| management | 90% | 92% | +2 |
| marketing | 90% | 95% | +5 |
| medical genetics | 92% | 92% | +0 |
| miscellaneous | 82% | 92% | +10 |
| moral disputes | 82% | 78% | -5 |
| moral scenarios | 65% | 52% | -12 |
| nutrition | 82% | 85% | +2 |
| philosophy | 98% | 95% | -2 |
| prehistory | 88% | 88% | +0 |
| professional accounting | 72% | 70% | -2 |
| professional law | 72% | 62% | -10 |
| professional medicine | 98% | 98% | +0 |
| professional psychology | 90% | 85% | -5 |
| public relations | 68% | 75% | +8 |
| security studies | 70% | 72% | +2 |
| sociology | 90% | 95% | +5 |
| us foreign policy | 92% | 95% | +2 |
| virology | 65% | 58% | -8 |
| world religions | 95% | 90% | -5 |
Also confirmed
| MTP speculative decoding (SGLang NEXTN) | preserved — cracked draft head agrees with the cracked model (coherent, ~2.4 accepted tokens/step) |
| Multimodal — image | ✅ working (shape/colour recognition, OCR) |
| Multimodal — video | ✅ working (motion/object description) |
| Coherence | no looping across code, math, reasoning, long-form (greedy) |
| Generation config | temperature 1.0, top_p 0.95, top_k 20 (stamped) |
Usage (SGLang, 2× DGX Spark TP2)
python -m sglang.launch_server \
--model-path dealignai/Qwen3.8-Flash-Next-ABLITERATED-NVFP4 \
--tp 2 --quantization modelopt_fp4 --fp4-gemm-backend flashinfer_cutlass \
--page-size 64 --mamba-scheduler-strategy extra_buffer --mamba-track-interval 64 \
--context-length 262144 --mem-fraction-static 0.80 --ple-offload-embedding
- Reasoning via
chat_template_kwargs:{"enable_thinking": true, "reasoning_effort": "xhigh"}(low|medium|xhigh). - NVFP4 — routed experts are 4-bit (NVFP4 W4A4); attention, shared experts, PLE n-gram table, vision tower and MTP remain higher precision. Serves on NVIDIA Blackwell (validated on 2× DGX Spark GB10).
- Serving note: for high-concurrency deployments, disable speculative decoding
(
--speculative-num-steps 0) for maximum decode stability; single-stream / low-concurrency serving runs NEXTN speculative decoding fine.
Disclaimer
This is an uncensored research artifact with safety refusals removed. You are solely responsible
for what you generate and for complying with all applicable laws. Provided as-is, without warranty.
Governed by the Qwen Community License 1.0 (see LICENSE).