Qwopus3.8-27B-Flash-V2, abliterated

Jackrong's GGUFs with the refusal direction projected out, stock llama.cpp loads them, no extra flag.

Source: Jackrong/Qwopus3.8-27B-Flash-V2-GGUF

How these were run

Same flags on a 4090, the file size is the weights, it is not the load, IQ4_XS at 262144 used 23098 MiB, Q3_K_L at 262144 did not fit with batch 4096 and MTP, so that command uses 32768 and loaded at 15738 MiB, MTP is inside the GGUF, no second draft file.

File	Quant	Size	4090	Context
`Qwopus3.8-27B-Flash-V2-IQ4_XS-abliterated.gguf`	IQ4_XS	14.36 GiB	23098 / 24564 MiB	262144
`Qwopus3.8-27B-Flash-V2-Q3_K_L-abliterated.gguf`	Q3_K_L	13.56 GiB	15738 / 24564 MiB	32768

llama-server -m Qwopus3.8-27B-Flash-V2-IQ4_XS-abliterated.gguf \
  --device CUDA0 --split-mode none --fit off -ngl 99 \
  -c 262144 -ctk q4_0 -ctv q4_0 \
  -b 4096 -ub 1024 --parallel 1 \
  --spec-type draft-mtp --spec-draft-n-max 3 --spec-draft-p-min 0.55 \
  --jinja --reasoning on --reasoning-effort low

llama-server -m Qwopus3.8-27B-Flash-V2-Q3_K_L-abliterated.gguf \
  --device CUDA0 --split-mode none --fit off -ngl 99 \
  -c 32768 -ctk q4_0 -ctv q4_0 \
  -b 4096 -ub 1024 --parallel 1 \
  --spec-type draft-mtp --spec-draft-n-max 3 --spec-draft-p-min 0.55 \
  --jinja --reasoning on --reasoning-effort low

Numbers

Runs used reasoning effort low, 100 prompts, lower is more open, strict is the canned refusal line, broader also counts softer refusals, think-off is 48 tokens, think-on scores the visible answer, up to 1024 tokens, one RTX 4090, WikiText-2, n_ctx 512, 580 chunks, HTML score is 0 to 100 on one demo page: browser, visual, features on screen, endurance.

	Thinking off, strict	Thinking off, broader	Thinking on, strict	Thinking on, broader	WikiText-2	Demo page	Score
IQ4_XS base	98	98	61	95	6.9287 ± 0.04415	ran, 13021 tok, PP 2180, TG 121	78
IQ4_XS abliterated	7	14	2	19	7.1203 ± 0.04595 (+2.8%)	ran, 16384 tok, PP 2090, TG 121	73
Q3_K_L base	98	98	56	92	7.2383 ± 0.04708	ran, 15387 tok, PP 1772, TG 110	69
Q3_K_L abliterated	8	18	1	10	7.3961 ± 0.04851 (+2.2%)	ran, 13564 tok, PP 1949, TG 114	76

Base rows are Jackrong's files, not this repo, WikiText-2 is a language-model score, not a coding score, on the Q3 bake, 3 think-on prompts never left the think block.

Edit

attn_out, ssm_out, ffn_down, token_embd, read from the Q8, projected, packed back into IQ4 or Q3, experts, output.weight, and MTP stay Jackrong's bytes, a weaker bake left the direction in place, pushing the live knob past the working point barely opened more prompts, and on Q3 the page broke.

Credits

Qwen team, Qwen3.8-27B, Qwen/Qwen3.8-27B
Jackrong, Qwopus3.8-27B-Flash-V2, the post-train and the GGUFs
Arditi, Obeso, Syed, Paleka, Panickssery, Gurnee, Nanda, 2024, Refusal is mediated by a single direction
Maxime Labonne, abliteration, the weight-space form
OrcaRouter, runtime reference while testing
llama.cpp, Georgi Gerganov and contributors, llama-cvector-generator measured r