Huihui-Qwen3.8-27B-abliterated-GGUF — AI-conn fork

This is a fork, not our model. All credit for the abliterated build goes to huihui-ai, and for the base model to Qwen (Qwen/Qwen3.8-27B). This fork exists so the fleet's copy cannot be altered or pulled upstream, and to carry the fit notes below. License: Apache-2.0 (per the source repo).

The build was made by abliteration (refusal-direction removal, using Sumandora's remove-refusals-with-transformers) — the method and tool are documented on the source card. Note from the same card: several sub-series (Swift / Saluki / Ternary variants) ablate only layers 22–52; the headline UD-DW and plain series are the main line.

Recommended file for a 20 GB card: Huihui-Qwen3.8-27B-abliterated-UD-DW-Q4_K_M.gguf (15.41 GiB).

Ollama pull:

ollama pull hf.co/AI-conn/Huihui-Qwen3.8-27B-abliterated-GGUF:Q4_K_M

Prime fit (measured budget, RX 7900 XT — 19.98 GiB VRAM, 32 GB RAM)

Fit math (weights + fp16 KV at 0.125 MiB/token + ~0.7 GiB runtime reserve):

UD-DW-Q4_K_M 15.41 GiB + 8k ctx (1.0 GiB) ≈ 17.1 GiB — fully resident
Same weights + 16k ctx (2.0 GiB) ≈ 18.1 GiB — fully resident

This is the only top-tier uncensored build in our audit with headroom for 16k context on a 20 GB card, which is why it is the fleet's primary brain. For comparison: any 32B-class Q4 build (stock or abliterated) leaves <1 GiB for KV + compute buffers on this card and partially CPU-offloads by arithmetic, whatever the card claims.