Research artifact. Abliteration removes content-refusal. It does not add capability. Do not ship this as a product default. The aligned repo remains the serving default.
Metric
Result
Quantization
ROCmFP4 103 FAST + 106 STRIX_LEAN; also Q6 114/116
The aligned card's finding still stands: the lever is MTP depth, not the ftype. llama.cpp's default --spec-draft-n-max 16 is about half the achievable throughput on this model. The knee is n=4. These uncensored files are the card's FAST (103) and STRIX_LEAN (106) from an abliterated BF16, plus Q6 AGENT/LEAN.
On the aligned family, FAST is dominated (same speed, worst PPL). Prefer 106 here unless you need the smallest file.
Q6 LEAN (ftype 116) is not the card LEAN. Card LEAN = 106.
Which file should I use?
Start with STRIX_LEAN (106) — the card-LEAN equivalent. Take Q6 AGENT (114) if you want protected heads. Do not default to FAST on this architecture.
Ryzen AI Max+ 395, ROCm 7.2.4, MTP --spec-draft-n-max 4, -fa on, ctx 8192, batch 1, temperature 0, thinking off so the 256-token budget is decode. Warm medians of 3.
Use the Q4_0 draft head, not Q8_0. Official mtp-Qwen3.8-27B-Q4_0.gguf works on this vocab (248k).
Flag
Why
--spec-draft-n-max 4
Default 16 lands on the wrong side of the curve.
--spec-draft-ngl 99
Without it the draft head can sit on CPU and the gain vanishes.
--jinja
Otherwise chat_template_kwargs are silently ignored.
-fit off
Autofit on iGPU can shrink context after an unload.
enable_thinking: falseorreasoning_effort: low
Default is xhigh. Small max_tokens then returns empty content. nonethrows.
⚠️ If the MTP draft command crashes on your build
The separate-model draft-mtp path has known bugs in the legacy charlie12345/ROCmFPX
line — reported on Windows 11 / gfx1151 / HIP SDK 7.2 (five stacked bugs, ending in an
h-row width mismatch in the draft's embedding buffer).
Fix: build the official repo instead — no patches needed.
git clone https://github.com/ROCmFPX/ROCmFPX.git
The MTP path was reworked there (unified n_embd_out row widths, t_h_nextn reset in
llm_graph_result::reset(), ctx_other wired centrally), which covers the whole chain.
Reported and verified on that configuration: 24–31 tok/s, coherent output, tool calling working.
Legacy-line patch: PR #109.
Linux builds on the legacy line are not known to be affected.
Speculative decoding (MTP)
Same recipe as the aligned card. Qwen3.8-27B has nextn_predict_layers = 1 as a separate draft GGUF, not in-model nextn.
Aligned MTP curve (STRIX, 8K): n=3 30.13 · n=4 30.30 (knee, accept 0.926) · n=5 27.52. An earlier sweep on uncensored Q6 LEAN (116) reproduced the same knee: n=3 24.96 · n=4 27.66 · n=5 25.27.
Uncensored findings
Content-refusal scoring on a 24 harmful / 12 harmless / 8 quality research set, greedy. Counts only. Refusal on the uncensored family was scored on Q6 AGENT (114); 4-bit 103/106 are the same checkpoint and should not restore refusals.
Model
Harmful 24
Harmless 12
Quality 8
Qwen3.8 aligned Q8 AGENT
23 refuse, 1 comply
11/12 ok (1 over-refuse)
6/8
Qwen3.8 uncensored Q6 AGENT (114)
23 comply, 1 broken
11/12 ok (1 over-refuse)
6/8
Muse aligned Q6 AGENT (114)
18 refuse, 6 comply
11/12 ok (1 over-refuse)
7/8
Muse uncensored STRIX_LEAN (106)
24 comply
12/12 ok (0 over-refuse)
7/8
Reading:
Aligned Qwen refuses ~all of this set. Abliterated Qwen complies ~all of it.
Quality is byte-identical as a score (6/8 both arms, same two fails). Abliteration here removed refusals without moving the smoke check.
One harmless over-refuse survived on both Qwen arms — the refusal classifier, not a unique unc defect.
Files
File
ftype
Size
Role
Qwen3.8-27B-Uncensored-ROCmFP4-FAST.gguf
103
13.33 GiB
smallest 4-bit; dominated on aligned PPL
Qwen3.8-27B-Uncensored-ROCmFP4-STRIX_LEAN.gguf
106
13.59 GiB
card-LEAN — recommended 4-bit
Qwen3.8-27B-Uncensored-Q6_0_ROCMFPX_AGENT.gguf
114
23.15 GiB
6-bit, Q8 head/attn
Qwen3.8-27B-Uncensored-Q6_0_ROCMFPX_LEAN.gguf
116
20.37 GiB
6-bit throughout
mtp-Qwen3.8-27B-Q4_0.gguf
—
1.56 GiB
MTP draft head — use this
mmproj-Qwen3.8-27B-Q8_0.gguf
—
0.59 GiB
vision projector (unmodified)
Six files. Aligned STRIX/FAST/LEAN are not in this repo.
No extra --tensor flags — matches the aligned recipe. Architecture is qwen35; no port required.
Architecture note (unchanged)
Dense hybrid attention, 64 layers, full_attention_interval = 4. KV is cheap. Prompt caching does not work on hybrid/recurrent memory in stock llama.cpp — budget full prefill every turn. MTP prompt-cache fix from the aligned repo is optional and not required to load these files.
Known issues
Vulkan/CUDA/CPU cannot load these files.
--spec-draft-n-max defaults to 16 — set 4.
reasoning_effort: "none" throws. Use enable_thinking: false.
Small max_tokens + thinking = empty content.
This is an uncensored research build. Do not deploy it as the public default.
Not yet measured
Test
Status
Perplexity vs aligned STRIX_LEAN 5.8871
❓ not re-run on unc
Tool-calling 7/7 on unc weights
❓ not re-run
Vision 4/4 spatial on unc
❓ projector reused
Independent reproduction
❓ none yet
License and attribution
Base model: Qwen team, Apache 2.0. MTP draft head redistributed with the Qwen GGUF companions. ROCmFP4 types: ROCmFPX. This repository is quantisation and measurement of an abliterated Qwen3.8-27B checkpoint.
See the aligned card for the MTP depth curve, tool-calling suite, and vision results.
