Qwen3.8-27B-Uncensored-Cyber

Cyber-specialized de-refusal of Qwen3.8-27B: fully open on the cyber/offensive-security domain while keeping reasoning, factual accuracy, and coherence intact. Vision tower and MTP speculative-decoding head are preserved (full multimodal, image-text-to-text).

Recipe (v2)

Built on the α=1.15 Aggressive base (recipe-v2 single-direction refusal ablation), then a residual-cyber peel: a cyber-pointed refusal direction (from the cyber-offensive training set vs broad harmless) removed by clean norm-preserving projection (β=1.0) applied only to the deeper layers (retain the first 4 layers, apply_from=4). Retaining the early feature-extraction layers is what preserves general capability — the lesson from the community's Qwen3 abliterations — so we reach 100% cyber-openness without wrecking the model.

Evaluation (bf16, larger-sample, Claude-judged; cyber = 100 held-out cyber-offensive prompts, regex refusal harness)

	cyber-open ↑	confab ↓	factual ↑	gsm8k ↑	degen ↓
Cyber (this model, v2)	100/100	0.867	1.00	0.80	0.00
previous Cyber build	93/100	1.00	0.933	0.825	0.00

The v2 recipe is more cyber-open (100 vs 93), less confabulating (0.87 vs 1.0), and more factually accurate (1.0 vs 0.93) than the previous Cyber, with equivalent reasoning and zero degeneration.

Quants

• -W4A16-AWQ — 4-bit weight AWQ (compressed-tensors), MTP head grafted
• -NVFP4 — NVFP4 (E2M1 4-bit / FP8 scales), MTP head grafted
• -GGUF — llama.cpp GGUF quants + vision mmproj + MTP head

Note

Uncensored / de-refused, tuned to fully answer cyber and offensive-security questions. Use responsibly and in compliance with applicable law.