The uncensored variant

qwen3.8-27b-abliterated is the same image with the huihui-ai abliteration applied. It is the same container: 1,255 of the 1,325 objects are byte-identical to the default image, and only the 70 matrices the abliteration changed are re-encoded. It does not refuse, so put the guardrails in the tool layer.

ignis-server model download qwen3.8-27b-abliterated
ignis-server --model-id qwen3.8-27b-abliterated   # found under --download-path, fetched when it is not there
make dev UNCENSORED=1   # the same, from a checkout

It is served as qwen3.8-27b-abliterated, so a client can tell it from the 27B. The pointing heads and the DFlash2 drafter load unchanged. They were fitted to the default weights, and the target still verifies every draft, so the drafter can only change the acceptance rate.