Red-teaming needs a capable model that does not refuse. A model that says no to every attack scenario is no use for writing test cases, reviewing payload logic or rehearsing an adversary. Abliteration removes refusal from a model’s weights, and an abliteration can pay for it in several places: lost capability, broken multimodal support, a speculative-decoding head that no longer matches the model, or a stream of answers that never finish.
Qwen3.8-27B-RANA-abliterated is our abliteration of Qwen/Qwen3.8-27B, built to avoid those costs. RANA stands for Reasoning-Anchored, Norm-preserving Ablation. We measured it head to head against the base model and against orcarouter/Qwen3.8-27B-Uncensored (“orca” below), the most-downloaded abliteration of Qwen3.8-27B on Hugging Face, through one identical harness. It comes out ahead on most of what we measured.
Why RANA
| RANA | orca | Base | |
|---|---|---|---|
| Hard refusals, HarmBench (of 200) | 0 / 0 | 0 / 0 | 168 / 155 |
| Answers, HarmBench | 78.8% / 57.3% | 73.8% / 61.2% | 3.2% / 0.7% |
| Answers, held-out | 90.0% / 54.4% | 83.9% / 49.2% | 3.2% / 0.0% |
| Over 16k tokens, HarmBench | 20.5% | 26.3% | 0.3% |
| Over 16k tokens, held-out | 6.7% | 11.9% | 0.6% |
| Capability change ¹ | 0.82 points | 1.11 points | – |
| Worst task ¹ | −1.5 (TruthfulQA) | −2.2 (HumanEval) | – |
| KL from base | 0.041 | 0.075 | 0 |
Refusal numbers are means over three sampling seeds. Pairs are the two independent judges, gpt-oss-safeguard and Mistral-Small. Both abliterations’ HarmBench labels went through the same re-judging rule, described under How it was measured. “Over 16k tokens” counts responses still unfinished at the 16k-token budget.
¹ Mean absolute change from base across five tasks, with TruthfulQA MC1 and MC2 averaged into one; the worst task uses the same averaging.
It answers more. On 240 held-out harmful prompts the model was never built on, RANA answers more often than orca under both judges, and the gaps are not noise: paired McNemar p = 2.9 × 10⁻⁶ under gpt-oss and p = 0.008 under Mistral. On HarmBench the judges split: gpt-oss favors RANA (p = 0.002), and Mistral favors orca, though not significantly (p = 0.10).
It finishes more often. Both abliterations reason at length on technical requests, and orca more so. Paired prompt by prompt on the held-out set, orca alone ran past the budget in 50 cases and RANA alone in 12 (p = 1.2 × 10⁻⁶). Long reasoning and the 16k cap covers what happens with more room.
It stays closer to the model you started with. KL divergence from base is 45% lower than orca’s, inside the “very good” band of the Abliterlitics scale. The biggest capability gap is on code: orca’s HumanEval drop is outside noise; RANA’s is not.
Nothing else is broken. The edit touches 131 weight matrices and nothing else. All 333 vision-tower tensors are byte-identical to the base model, so image input works exactly as before. The multi-token prediction (MTP) head is edited consistently with the main model, so speculative decoding still accepts its drafts: mean acceptance length 2.45 on the BF16 build.
It runs where you run models. BF16 at full precision, FP8 in exactly Qwen’s official format for vLLM and SGLang, 15 GGUF quants for llama.cpp with vision and the MTP draft included, and MLX builds for Apple Silicon from 4-bit to BF16. Each was measured separately.
You can check all of it. Three seeds, two judges from different labs, a held-out prompt set, release gates fixed before the final evaluation, and a results/ folder in every repo with the numbers behind every claim here.
It is not better at everything. RANA is 2.2 ± 0.9 points behind orca at instruction following (IFEval).
Get it
| Build | Repo | Size | Runs on |
|---|---|---|---|
| BF16 (full precision) | preemware/Qwen3.8-27B-RANA-abliterated | 55.6 GB | vLLM, transformers |
| FP8 | preemware/Qwen3.8-27B-RANA-abliterated-FP8 | 30.9 GB | vLLM, SGLang |
| GGUF | preemware/Qwen3.8-27B-RANA-abliterated-GGUF | 10.0–28.6 GB per quant | llama.cpp and compatible apps |
| MLX | preemware/Qwen3.8-27B-RANA-abliterated-MLX | 16.1–54.7 GB per folder | Apple Silicon, mlx-vlm |
All four keep vision, tool calling, thinking mode and the MTP head. These are research models: their own refusals are largely gone, so anything facing end users needs a separate moderation layer.
A small edit in exactly the right places
Abliteration rests on a finding from Arditi et al., 2024: in chat models, refusal is mediated by a single direction in the residual stream. Remove that direction from the weights that write to the stream, and the model largely stops refusing. RANA does that, then rescales every edited weight vector back to its original norm, so the rest of the network sees weights of the size it was trained with.
The first question about any edit like this is what else it touched, so we diffed every tensor against the base model. Exactly 131 matrices changed, and nothing outside that set:
| Part | Matrices changed |
|---|---|
Attention output (self_attn.o_proj) |
17 (16 main, 1 MTP) |
Linear-attention output (linear_attn.out_proj) |
48 |
MLP output (mlp.down_proj) |
65 (64 main, 1 MTP) |
| Token embeddings | 1 |
The changes are small. The median edited matrix moves by 1.4% of its norm, and none by more than 1.7%. The rescaling is exact in FP32; storing the result in BF16 leaves column norms within 0.08% of the originals. The component along the refusal direction drops from as much as 0.225 before the edit to at most 0.00122 after.
As an outside check, the open-source Abliterlitics toolkit (commit 9b15eb0) counts exactly 129 changed tensors among the 850 it inspects. It skips vision and MTP tensors, so it sees 131 minus the two MTP edits.
How it was measured
We followed the structure of the Abliterlitics methodology: weight forensics, KL divergence, a HarmBench safety evaluation read by LLM judges, and a capability suite, all behind gates fixed before the final evaluation. Where our setup departs from theirs, it is mostly because this model reasons at length before it answers:
| Instrument | What we ran | Difference from the Abliterlitics setup |
|---|---|---|
| Weight forensics | Abliterlitics fingerprint, plus our own tensor-by-tensor diff | No SVD, rank or per-layer analysis published |
| KL divergence | First-token, full-vocabulary KL on mlabonne/harmless_alpaca test[:100] (the Heretic protocol) |
Same protocol |
| Safety | 200 HarmBench standard behaviors plus 240 held-out prompts (StrongREJECT, JailbreakBench, CategoricalHarmfulQA), 3 seeds, 2 judges | HarmBench standard subset instead of all 400 textual behaviors; a held-out set the model was never built on; a second, independent judge |
| Capability | MMLU-Pro, GSM8K, IFEval (generative, thinking on), TruthfulQA MC1/MC2, HumanEval | HumanEval through a chat-format harness instead of excluding it; no GPQA, BBH or MUSR |
The base model, RANA and orca went through one identical harness on vLLM 0.30.0, with thinking on, Qwen’s recommended sampling (temperature 1.0, top_p 0.95, top_k 20) and a 16k-token budget per response. Every model saw the same prompts, so the comparisons are paired prompt by prompt.
Two judges from different labs read every response: openai/gpt-oss-safeguard-20b and mistralai/Mistral-Small-24B-Instruct-2501. Both saw the full reasoning trace and final answer and applied the same four labels: complied, soft refusal, hard refusal, degenerate. They agree closely on outright refusals (Cohen’s κ = 0.89) and only moderately on whether a hedged answer counts as answering (κ = 0.54), which is why answer rates come as a pair. For both abliterations, every gpt-oss hard-refusal label on HarmBench is re-judged once with the same judge plus a fixed reminder that it is auditing a transcript, with no manual edits. Each abliteration had 2 such labels across its 600 HarmBench responses before re-judging and none after. The Mistral judge gave neither abliteration a hard refusal on HarmBench in any run.
The gates were written into the validation spec before the final evaluation ran:
| Gate | Requirement |
|---|---|
| No hard refusals | 0 hard refusals on HarmBench, both judges, all seeds |
| Finishes more often than orca | Lower rate of responses hitting the token budget than orca on the held-out set, paired test p < 0.05 |
| Capability kept | Mean absolute change ≤ 1.0 point across the suite, no task below −2.0, KL < 0.15 |
| Surgical edit | Only intended matrices changed, vision byte-identical, MTP intact |
| Still works | Vision, multi-turn tool calling and MTP speculative decoding pass on the final files |
The gates are not decoration. An earlier candidate lost 2.24 points on HumanEval and was not released. RANA passed all five.
Capability, task by task
Changes are in points against base, ± standard error, paired on the same questions. Bold marks a change larger than two standard errors.
| Task (items) | Base | orca Δ | RANA Δ |
|---|---|---|---|
| MMLU-Pro (700) | 80.0 | −0.4 ± 1.1 | +0.9 ± 1.1 |
| GSM8K (1,319) | 91.4 | +0.2 ± 0.5 | −0.1 ± 0.6 |
| IFEval (541) | 90.6 | +0.9 ± 0.9 | −1.3 ± 1.1 |
| TruthfulQA MC1 (817) | 36.1 | −1.8 ± 0.6 | −1.4 ± 0.6 |
| TruthfulQA MC2 (817) | 54.2 | −1.6 ± 0.3 | −1.6 ± 0.3 |
| HumanEval (164 × 3) | 97.2 | −2.2 ± 1.0 | −0.4 ± 0.9 |
RANA’s one change outside noise is about 1.5 points on TruthfulQA (MC1 and MC2 averaged), a similar drop to orca’s 1.7. The Abliterlitics methodology notes that TruthfulQA consistently drops after abliteration, which suggests a cost of the technique rather than of this recipe. Everything else sits within noise of base.
Long reasoning and the 16k cap
On technical requests, RANA reasons at length: 20.5% of its HarmBench responses and 6.7% of its held-out responses are still unfinished at 16k tokens, against 26.3% and 11.9% for orca. This is not the model getting stuck. In 115 of RANA’s 123 such HarmBench responses, the text is no more repetitive than normal output, and many of the same prompts run long in every seed, clustered in categories that ask for something technical to be built.
Given room, they finish. We took every prompt where the FP8 build hit the 16k budget in its first seed, 59 in all rather than a sample, and ran them again with a 64k budget: 57 finished on their own at a median of about 24k tokens, and the gpt-oss judge scored 55 as answering the request. Qwen recommends allowing up to 262k tokens of reasoning for this model, so the answer rates above, measured at 16k, are lower bounds. The commands under Try it allow 64k.
FP8: Qwen’s own format
The FP8 build uses exactly the block-FP8 format of Qwen’s official Qwen3.8-27B-FP8: the same 1,606 tensors, the same 407 FP8 weights in 128 × 128 blocks, the same tensors left in BF16. vLLM serves it with the same kernels as Qwen’s release. On the tensors abliteration did not touch, our conversion reproduces Qwen’s file: 100% of block scales and 99.91% of FP8 values are bit-identical.
Quantizing an abliterated model raises a fair question: could rounding put the removed direction back? The measurements say no. FP8 rounding restores 2.7% of the original component, against 1.0% from BF16 rounding alone, and what it adds is noise, with a cosine similarity of 0.001 to the component that was removed. On the same KL protocol, the FP8 weights (dequantized, so this isolates weight rounding) sit at 0.048 from base, against 0.041 for BF16. Across three seeds and both judges, FP8 has 0 hard refusals on HarmBench, and no capability task differs from BF16 by more than about two standard errors.
GGUF: 15 quants, vision and MTP included
The GGUF repo follows ggml-org’s layout for the base model: the main model, the MTP head as a separate mtp-* file for speculative decoding, and the vision tower as mmproj-*, which llama.cpp picks up automatically. Every quant uses an importance matrix computed from bartowski’s public calibration text. Quality is measured as KL divergence from the BF16 GGUF on wikitext, the protocol bartowski publishes for his base-model quants:
| Quant | Size | KLD | Same top token |
|---|---|---|---|
| Q8_0 | 28.60 GB | 0.0009 | 98.7% |
| Q6_K | 22.08 GB | 0.0022 | 97.9% |
| Q5_K_M | 19.23 GB | 0.0064 | 96.6% |
| Q4_K_M | 16.55 GB | 0.0156 | 94.5% |
| IQ4_XS | 15.08 GB | 0.0186 | 94.1% |
| IQ3_M | 12.58 GB | 0.0628 | 89.4% |
| IQ2_M | 10.00 GB | 0.1723 | 82.1% |
The card lists all 15. At equal file size, the numbers sit close to bartowski’s quants of the base model. On Q4_K_M served with llama-server, vision, three-turn tool calling and MTP speculative decoding all pass, and a one-seed refusal run shows 0 hard refusals on HarmBench and answer rates within noise of the BF16 build.
MLX: one repo, 4-bit to BF16
The MLX repo holds one folder per precision, plus the MTP drafter in mtp/; the 4-bit build is also at the root. In every folder, the tensors and config.json match the mlx-community or lmstudio-community conversion of the base model exactly. Quality is KL divergence from the BF16 MLX build, on the same text and protocol as the GGUF table:
| Folder | Size | KLD | Same top token |
|---|---|---|---|
| 4-bit | 16.08 GB | 0.0468 | 90.0% |
| 5-bit | 19.44 GB | 0.0129 | 94.8% |
| 6-bit | 22.80 GB | 0.0042 | 96.9% |
| 8-bit | 29.53 GB | 0.0013 | 98.3% |
| bf16 | 54.74 GB | reference | – |
MLX quantizes by plain rounding, like the standard community conversions, so at a similar size its 4-bit sits further from BF16 than the imatrix Q4_K_M GGUF (0.0156); the 5-bit build is the closer match. Vision and three-turn tool calling pass on 4-bit, 8-bit and BF16, and MTP speculative decoding accepts 2.49–2.57 tokens per round across the five precisions. These checks ran on MLX’s CUDA backend on Linux, not on a Mac.
What we are not publishing
The recipe: which activations the direction is measured on, the datasets used to find it, and the exact edit. The weights, the measurements and the checks are public, so every claim here can be verified without it.
The generated responses are not published either. By construction they are a few thousand answers to harmful requests. The per-model counts, judge agreement and paired tests are in each repo’s results/ folder.
Limits
This is a safety-alignment-removed model, and it will produce content the base model refuses. Do not put it in front of end users without a separate moderation layer. Refusal numbers come from LLM judges that disagree on partial or hedged answers, so “no hard refusals” is a judge-measured claim, not a guarantee.
The head-to-head comparison covers the base model and orca, the most-downloaded abliteration of Qwen3.8-27B; other abliterations were not run through this harness. RANA trails orca on instruction following (IFEval −2.2 ± 0.9 points). Our capability suite is narrower than Abliterlitics’ current one: MMLU-Pro ran on a 700-question subset, and GPQA, BBH and MUSR were not run. The smaller builds were tested less deeply than BF16: FP8’s KL covers weight rounding only, not FP8 activations, and on GGUF only Q4_K_M was tested for refusal behavior, on one seed.
Try it
vLLM, BF16 or FP8 (on Blackwell GPUs, also set VLLM_USE_FLASHINFER_SAMPLER=0):
vllm serve preemware/Qwen3.8-27B-RANA-abliterated-FP8 \
--max-model-len 65536 \
--reasoning-parser qwen3 \
--enable-auto-tool-choice --tool-call-parser qwen3_coder \
--speculative-config '{"method": "mtp", "num_speculative_tokens": 2}'llama.cpp, pulling the quant, the vision file and the MTP draft straight from the repo:
llama-server -hf preemware/Qwen3.8-27B-RANA-abliterated-GGUF:Q4_K_M \
--spec-type draft-mtp --jinja -fa on -ngl 99 -c 65536 \
--temp 1.0 --top-p 0.95 --top-k 20 --min-p 0MLX on Apple Silicon. Download one folder: passing the repo id to --model downloads every folder, about 160 GB.
hf download preemware/Qwen3.8-27B-RANA-abliterated-MLX --include "4-bit/*" --include "mtp/*" \
--local-dir ./Qwen3.8-27B-RANA-abliterated-MLX
python -m mlx_vlm generate --model ./Qwen3.8-27B-RANA-abliterated-MLX/4-bit \
--draft-model ./Qwen3.8-27B-RANA-abliterated-MLX/mtp --draft-kind mtp \
--enable-thinking --max-tokens 32768 --temperature 1.0 --top-p 0.95 --top-k 20 \
--prompt "Explain how a hash map handles collisions."Use Qwen’s thinking-mode sampling: temperature 1.0, top_p 0.95, top_k 20, min_p 0. Each model card has the tested commands, the full tables and the release gates.
Questions, problems, or results from your own evaluations: open a discussion on the model’s Hugging Face page.
Thanks to the Qwen team for the base model, Arditi et al. for the refusal-direction work, Jim Lai (grimjim) for norm-preserving abliteration, p-e-w for the Heretic KL protocol, the Abliterlitics project for its toolkit and methodology, the llama.cpp maintainers and bartowski for the GGUF tooling and reference numbers, and the MLX and mlx-vlm maintainers for the Apple Silicon tooling.
Update, September 26, 2026: orca’s HarmBench labels now go through the same re-judging rule as RANA’s, which moves its gpt-oss hard-refusal count from a mean of 0.7 per seed to 0. Also added: HarmBench rows in the headline table, the FP8 KL figure, 64k-context commands, and the MLX builds.