MTP / nextn is omitted (--no-mtp). Speculative decode does not make the model smarter; the extra head steals KV on 12–24 GB cards. Need vision? Pair a separate mmproj. Need MTP? Use another pack.
The source checkpoint had the refusal direction removed (abliteration). These GGUFs inherit that behavior.
Chat template
Official 3.8 jinja wraps every assistant turn in <think>…</think> even when reasoning is empty, then opens another <think> on generate. That truncates multi-turn agents.
These GGUFs bake a fixed template. Use --jinja. A standalone chat_template.jinja ships in the repo if an older copy is still on disk.
Q6_K is the largest file that still full-offloads 24GB Turing. Q8_0 does not (-ngl 99 will not fit).
Intended use
Local llama.cpp serving of the uncensored 27B trunk: research, red-team, and unfiltered generation in a setting you control.
Out of scope: treating this as official Qwen or as a drop-in for vcruz305/Qwen3.8-27B-GGUF; deploying to end users without your own filters; any use that breaks Apache-2.0 or the law.
Bias, risks, limitations
Safety alignment was removed at the source. The model will answer requests the official 27B would refuse. It still carries the bias and failure modes of Qwen3.8-27B, plus K-quant error. These files are language-only (no vision tower, no MTP).
