Qwen3.8-27B-Uncensored — Q38mix v6 (edit-preservation-aware GGUF)

A reasoned GGUF quantization of orcarouter/Qwen3.8-27B-Uncensored (the rank-1 "OrcaRouter" abliterated Qwen3.8-27B, a hybrid gated-delta-net + attention architecture with a native MTP head). 19.32 GB, 5.655 bpw. Fits an RTX 5090 (32 GB) with headroom for KV cache / long context.

What makes it different: to our knowledge it is the first GGUF quant of this model that is aware of the abliteration edit itself and protects the tensors carrying it, while being tuned per-tensor against a true BF16 reference (not a Q8 proxy).

Method (every choice measured)

1. Custom importance matrix calibrated on this model: 861k tokens (agentic-coding + code + chat + Italian
   • refusal-contrast), Qwen3 chat-template rendered, 125 chunks @ ctx 2048 with --process-output.
2. Per-tensor base allocation = Unsloth Dynamic V3 (extracted from unsloth/Qwen3.8-27B-GGUF UD-Q4_K_XL) — full credit to Unsloth for the base map.
3. Survival-guided edit protection: the 131 tensors carrying the shared rank-1 abliteration direction (recovered first-hand, cosine 0.99999 across layers) are protected.
4. Measured reallocation: all ffn_down → Q6_K, attn_output → Q6_K, token_embd → Q6_K, driven by a ΔKL-attribution sweep and validated against a BF16 reference. ssm_alpha/beta kept Q8_0; conv1d/dt/a and norms F32 (GDN control path, incompressible). lm_head/output not inflated (least sensitive — placebo).

Benchmarks — against a BF16 reference (the bias-free objective)

KL divergence vs the true Uncensored BF16 (held-out multi-domain corpus, 40×512 chunks). Reference is BF16, so the comparison does not favour any build's Q8-kept tensors.

Honest reading:

• Beats the publisher's own OrcaRouter Q5_K_M on every measured axis — mean KL −14.5 %, tails −17 %/−19 %, and it is smaller. This is a fair BF16-referenced comparison.
• On par with the Unsloth method (udprotected): v6 edges mean KL and tails but is +0.23 GB larger, and udprotected wins same-top-1 by 0.04 pp (noise). At equal size the two are a tie. Do not read this as "beats Unsloth" — the differentiator over the Unsloth method is the edit protection, not the allocation.
• Edit-survival (structural preservation of the abliteration direction): min ≈ 0.95, on par with OrcaRouter Q5 and far above an unprotected Unsloth-style quant (~0.86).

Trajectory fidelity (antirez ds4: greedy-LCP vs the Q8 reference's own greedy path, n=46)

v6 is co-leader on greedy-LCP (tied with OrcaRouter Q5 within noise) and best on exact-match, ahead of the Unsloth method — i.e. it is strong on both the KL metric and the trajectory metric, which usually trade off. (n=46; the ~0.6-LCP gaps are within the sign-test noise band — read as "competitive/co-leading", not "decisively best".)

Honest caveats (read before relying on this)

• Reference is BF16 for the table above (the correct objective). An earlier Q8-referenced number for this family was self-favourable and is not used here.
• Behavioural refusal-rate (measured, full-window): on 25 AdvBench harmful + 15 Alpaca benign prompts, n_predict=640 with the refusal judged on the answer after </think> (24/25 harmful outputs closed the think block inside the window — a valid measurement, not the 128-token truncation the earlier bench suffered): harmful-refusal 0.0, benign-refusal 0.067 — identical to the Q8 baseline. The abliteration is preserved behaviourally; OrcaRouter Q5, by contrast, restored refusals (0.017 / 0.100).
• Calibration/eval overlap: ~43 % of the KL-eval tokens share a source distribution with the calibration corpus (documents disjoint, distribution not) — this favours any build using this imatrix over externally-built quants. The numbers are honest but this is disclosed.
• Built from the Q8_0 source with --allow-requantize (a BF16-sourced rebuild is in progress and will supersede this if measurably better).

Provenance

• GGUF sha256: 0e6f25cf9d203e0c75b9cb814e47346edc0d80b293d6a151287049da22ad4e04
• Engine: llama.cpp b10470 · quantized_from: Uncensored Q8_0 (--allow-requantize)
• Calibration: 861,596 tokens, 7 buckets · imatrix: 125 chunks @ ctx 2048, --process-output
• Full per-tensor map, imatrix, and calibration recipe: see provenance.json and the accompanying files.

Variants in this repo

• …Q38mix-v6.gguf (19.32 GB) — the flagship above (recommended).
• …Q38mix-v8-19.1G.gguf (19.14 GB, 5.60 bpw) — a slightly smaller variant (ffn_gate/up → IQ4_XS). BF16-ref KL 0.008593, same-top-1 96.35 % (best of the set), tail (kld99) 0.098 — i.e. ≈v6 quality within noise, −0.18 GB, with only the tail marginally worse. This is the mainline quality-per-byte frontier: a size-matched reallocation (v6b) does not beat v6, and a real ~2 GB reduction is only reachable with the ik_llama.cpp trellis fork (not servable in mainline llama.cpp / ollama / LM Studio). Full study: the project's docs/quantization/ and research/quant-obsessive/10-shrink-v6-executable.md.

Intended use

A high-fidelity local GGUF of an existing public abliterated model, for single-user coding-agent workloads on a 24–32 GB consumer GPU. It is a quantization of a model that is already publicly available; it adds no capability beyond faithfully preserving that model at smaller size.