Qwen3.8-27B-Uncensored-Aggressive (bf16)

Uncensored (refusal-suppressed) build of Qwen3.8-27B (dense VLM, qwen3_5 arch), for NVIDIA V100 (SM70) under 1Cat-vLLM.

Abliteration

Heretic KL-optimized directional abliteration (aggressive/broad profile). Pareto trial 192: refusals reduced from 99/100 to 14/100 on mlabonne/harmful_behaviors, KL divergence 0.106 vs the stock base. Sanity (bf16): 0/5 harmful prompts refused, coherence intact. Direction computed and refusals evaluated in non-thinking mode (/no_think). Vision tower and MTP head untouched.

Format

bf16 merged weights. MTP head (mtp.*) carried verbatim from the base.

Serve (1Cat-vLLM, 2x V100, TP2)

--kv-cache-dtype fp8_e5m2, MTP speculative decoding, --gpu-memory-utilization tuned to KV/context budget.

Responsible use

Research model with safety refusals removed. Use lawfully and responsibly.