dealignai/Qwen3.8-Flash-Next-ABLITERATED-FP8

Qwen3.8-Flash-Next — ABLITERATED (FP8)

Abliterated (refusal-removed) build of Qwen/Qwen3.8-Flash-Next in official FP8.
Reasoning (low / medium / xhigh), MTP speculative decoding, and full multimodality (image + video) preserved.

by dealignai

---

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 — nothing special is required at inference time. 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	low	100 %
greedy	xhigh	99.6 %
greedy	off	97.1 %

With reasoning on (low / xhigh) it is essentially complete (≈ 100 %). The small residual appears only with both reasoning off and greedy (the hardest possible setting) — and even that is 97 %. Turn on reasoning or use any non-zero temperature and it complies.

Per-category (real-harm compliance, greedy)

Category	reasoning off	reasoning low	reasoning xhigh
chemical / biological	100 %	100 %	100 %
cybercrime / intrusion	100 %	100 %	100 %
illegal	98 %	100 %	100 %
misinformation	98 %	100 %	98 %
harmful	94 %	100 %	100 %
harassment / bullying	81 %	100 %	100 %

Every category reaches 100 % with reasoning on. The only dip is harassment under reasoning-off.

Capability retained — MMLU (identical harness, base vs this model · 40 Q/subject · 2 280 Q)

Overall: 86.36 % → 83.86 % (-2.50 pp) — capability preserved (within 3 pp of base).

Per-subject breakdown — all 57 MMLU subjects

Subject	Base	This model	Δ
abstract algebra	72%	57%	-15
anatomy	88%	80%	-7
astronomy	98%	98%	+0
business ethics	82%	82%	+0
clinical knowledge	98%	90%	-7
college biology	100%	100%	+0
college chemistry	60%	55%	-5
college computer science	85%	80%	-5
college mathematics	72%	65%	-7
college medicine	88%	85%	-3
college physics	72%	70%	-3
computer security	88%	90%	+3
conceptual physics	92%	92%	+0
econometrics	82%	82%	+0
electrical engineering	85%	80%	-5
elementary mathematics	92%	90%	-3
formal logic	70%	70%	+0
global facts	57%	60%	+3
high school biology	95%	95%	+0
high school chemistry	88%	92%	+5
high school computer science	90%	90%	+0
high school european history	88%	90%	+3
high school geography	95%	98%	+3
high school government and politics	98%	98%	+0
high school macroeconomics	92%	90%	-3
high school mathematics	52%	55%	+3
high school microeconomics	100%	98%	-3
high school physics	90%	88%	-3
high school psychology	98%	95%	-3
high school statistics	88%	85%	-3
high school us history	95%	95%	+0
high school world history	95%	95%	+0
human aging	85%	82%	-3
human sexuality	92%	90%	-3
international law	88%	82%	-5
jurisprudence	98%	95%	-3
logical fallacies	95%	90%	-5
machine learning	80%	72%	-8
management	98%	98%	+0
marketing	92%	90%	-3
medical genetics	95%	95%	+0
miscellaneous	95%	90%	-5
moral disputes	78%	75%	-3
moral scenarios	80%	70%	-10
nutrition	92%	88%	-5
philosophy	80%	82%	+2
prehistory	98%	95%	-3
professional accounting	78%	68%	-10
professional law	72%	65%	-7
professional medicine	98%	98%	+0
professional psychology	82%	75%	-7
public relations	80%	75%	-5
security studies	90%	90%	+0
sociology	92%	88%	-5
us foreign policy	95%	95%	+0
virology	52%	48%	-5
world religions	92%	95%	+2
Overall	86.36 %	83.86 %	-2.50

Also confirmed

MTP speculative decoding (vLLM qwen3_8_flash_next_mtp)	preserved — ~81 % draft acceptance (≈ 1.8× draft efficiency)
Multimodal	image + video both working
Coherence	no looping across code, math, reasoning, long-form (greedy)
Generation config	temperature 1.0, top_p 0.95, top_k 20 (stamped)

Usage (vLLM)

from vllm import LLM, SamplingParams
llm = LLM(model="dealignai/Qwen3.8-Flash-Next-ABLITERATED-FP8",
          tensor_parallel_size=2, trust_remote_code=True)
# reasoning via chat_template_kwargs: {"enable_thinking": True, "reasoning_effort": "xhigh"}  # low | medium | xhigh

• PLE n-gram table is CPU-offloaded at runtime: set VLLM_PLE_CPU_OFFLOAD=1.
• MTP speculative decoding: speculative_config={"method": "qwen3_8_flash_next_mtp", "num_speculative_tokens": 1}.
• FP8 — serves on vLLM (Hopper / Blackwell; runs on 2× DGX Spark).

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).