license: other
license_name: qwen-community-1.0
license_link: LICENSE
base_model: ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-GGUF
base_model_relation: finetune
tags:
- uncensored
- abliterated
- gguf
- strata
Qwen3.8-Flash-Next GSQ-RCO IQ3_S — uncensored, layers 6–41
We present L6-41, an uncensored version of ISTA-DASLab's GSQ-RCO IQ3_S of Qwen3.8-Flash-Next. It differs from our earlier release, q8g, in two ways: all 156 attention tensors were replaced with Q8_0 tensors made from the official BF16 weights, and the refusal direction was removed only from layers 6–41 (108 matrices, against 135 in layers 0–44 for q8g).
On HarmBench, L6-41 complied with 207 of 240 prompts, against 5 for the same file without the refusal edit and 204 for q8g. The KL divergence of its next-token probabilities from the original GSQ-RCO file was 0.0333 (q8g: 0.0367). Its scores on BigCodeBench-Hard, IFEval and 100 text-only questions of Humanity's Last Exam did not differ significantly from those of the unedited file or of q8g.
1. What changed from q8g
q8g removed one refusal direction per layer from the 135 matrices that write to the residual in layers 0–44, and kept the rest of the GSQ-RCO file as it was (q8g card). For L6-41:
- Attention at Q8_0. We first made a file in which all 156 attention tensors of the 48 layers (Q4_K–Q6_K and IQ4_XS in the original) were replaced with Q8_0 tensors quantized from the official BF16 weights of Qwen/Qwen3.8-Flash-Next. Below, this file is called the chimera. It has no refusal edit, and it complied with only 5 of 240 HarmBench prompts.
- Fewer layers edited. We searched which layer ranges are enough to stop refusals (Section 2) and edited only layers 6–41 of the chimera.
- Same kind of direction. One refusal direction per layer, taken from the same 400 prompts as q8g, but recorded on the chimera. In layers 4–44, its cosine similarity with q8g's direction was at least 0.9993.
The result is 0.90 GB larger than the original file (Table 1).

Figure 1. L6-41 edits layers 6–41; q8g edited layers 0–44. Every attention tensor of the 48 layers is Q8_0 made from BF16. In layers 6–41, the refusal direction was removed from the 3 matrices that write to the residual.
Table 1. Tensors that differ from the original GSQ-RCO file in shard 1.
| Tensor family | Layers | Count | Type in the original | Type in L6-41 | Refusal direction removed | Size |
|---|---|---|---|---|---|---|
attn_qkv, attn_gate, ssm_out |
linear-attention layers (36) | 36 each | Q4_K / Q5_K / Q6_K / IQ4_XS | Q8_0 | ssm_out in layers 6–41 (27) |
+0.672 GB |
attn_q, attn_k, attn_v, attn_output |
full-attention layers 3, 7, …, 47 (12) | 12 each | Q4_K / Q5_K / Q6_K / IQ4_XS | Q8_0 | attn_output in layers 7–39 (9) |
+0.197 GB |
ffn_down_shexp |
6–41 | 36 | IQ4_NL | Q8_0 | all 36 | +0.029 GB |
ffn_down_exps |
6–41 | 36 | IQ4_NL×28 / Q2_0×8 | unchanged formats, codes re-chosen | all 36 | ±0 |
| Total | 228 | 108 | +0.90 GB |
The other 995 tensors of shard 1, and all of shard 2, are identical to the original's.
Method in full
1.1 Starting file and tools
- Model:
IQ3_S/Qwen3.8-Flash-Next-GSQ-RCO-IQ3_S-00001-of-00002.gguffrom ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-GGUF (54,817,524,224 bytes). Shard 2 was not changed. - Attention tensors: the official BF16 weights of Qwen/Qwen3.8-Flash-Next, quantized to Q8_0. The largest relative error of the Q8_0 tensors against BF16 was 0.0087.
- Recording activations: Strata 0.2.10 with a patch that writes the residual after each layer.
- Weight edits: Python, gguf-py from llama.cpp, PyTorch on one RTX 5090.
- Recording inputs to the routed experts: llama.cpp with a tool that records the inputs to each routed expert's down matrix (as for q8g).
- Evaluation: Strata 0.2.15 (Section 3).
1.2 Refusal direction
- Prompts: the first 200 train prompts of mlabonne/harmful_behaviors (to be refused) and of mlabonne/harmless_alpaca (ordinary), each as one user turn in the chat template. These are the same 400 prompts as for q8g.
- The chimera read each prompt. After each of the 48 layers, we took the residual as the mean of the 4 hyper-connection streams, and averaged it over all tokens of the prompt.
- For each layer, the refusal direction is the mean over the 200 refused prompts minus the mean over the 200 ordinary prompts, normalized to length 1. No orthogonalization and no masked dimensions.
1.3 Matrices edited
- In each of layers 6–41, the 3 matrices that write to the residual: the attention output (
ssm_outorattn_output), the shared-expert down (ffn_down_shexp) and the routed-expert down (ffn_down_exps, 512 experts). 36 layers × 3 = 108 matrices. - From each, we removed the component along that layer's direction u: W' = W − u uᵀW.
1.4 How each matrix was saved
- Attention output (36). Dequantized the chimera's Q8_0 weights, computed W', and quantized to Q8_0. Where more than 2% of the component along u was left after rounding (all 36 matrices), the codes were adjusted, with the scales fixed, to bring it down.
- Shared-expert down (36). Dequantized the original IQ4_NL weights, computed W', and quantized to Q8_0, with the same adjustment where more than 2% was left (28 of 36).
- Routed-expert down (36 × 512 experts). Kept the original formats and scales, and re-chose only the codes with GPTQ toward W', constrained against u, as for q8g (Section 2.5 of the q8g card): the chimera read the same 8 coding-agent transcripts as for q8g, and the inputs to each expert were recorded at every 4th token; damping 0.01, act order, blocks of 128. For experts with fewer than 64 recorded tokens (6–67 of 512 per layer), direct code changes were used instead.
1.5 Result
- Component along u left after saving (before = 1): at most 0.0007 for attention output, 0.0196 for shared-expert down, 0.0030 for routed-expert down (6 of 512 experts sampled per layer).
- Output shard 1: 55,716,585,984 bytes, SHA-256
3e784d5b9d32c8033292d3f562e17dddde177bb52c1f6fe6a0243013e290a1bb. Shard 2 is identical to the original's.
2. How we chose layers 6–41
2.1 Layers needed to stop refusals
Before editing weights, we ran the chimera with the direction removed during generation, over different layer ranges: in each layer of the range, Strata removed the component along that layer's direction from the outputs of the same 3 matrices, which is what the weight edit does. We counted refusals on the first 50 refused prompts (1 run each, temperature 1.0). A local Qwen3.8 27B labeled each answer as refused, hedged, partial or complied. These 50 prompts are among the 200 used to make the direction, so this search only compares layer ranges; HarmBench (Section 3.1) is the test on unseen prompts.
No single block of about 10 layers was enough (refused: 46, 40, 43, 32 and 49 of 50 for layers 0–9, 10–18, 19–28, 29–37, 38–47). At the 3 positions tested with 16 layers (14–29, 15–30, 16–31), 10–15 prompts were still refused. Some 18-layer ranges brought this down to 3–4 (14–31, 16–33), but not all (10–27: 18). Ranges of 24–26 layers around layers 10–35 complied about as often as the 45 layers of q8g (Table 2).

Figure 2. Layer ranges: the best 16-layer range left 10 refusals; the best ranges of 22 layers or more left none. Left: layers with the direction removed during generation. Right: answers to 50 refused prompts, labeled by a local Qwen3.8 27B. 1 run each.
Table 2. Best range for each number of layers (direction removed during generation), and the range we chose. KL: from the chimera, 4 coding-agent transcripts (Section 3.2).
| Layers | Count | Refused / 50 | Hedged | Complied | KL |
|---|---|---|---|---|---|
| 16–31 | 16 | 10 | 24 | 16 | 0.0106 |
| 16–33 | 18 | 4 | 12 | 34 | 0.0111 |
| 13–32 | 20 | 2 | 10 | 38 | 0.0138 |
| 14–35 | 22 | 0 | 9 | 41 | 0.0138 |
| 12–35 | 24 | 0 | 7 | 43 | 0.0150 |
| 10–35 | 26 | 0 | 5 | 45 | 0.0164 |
| 6–41 | 36 | 0 | 2 | 47 (+1 partial) | 0.0214 |
| 0–44 | 45 | 0 | 4 | 46 | 0.0366 |
| q8g (weights edited) | 45 | 0 | 5 | 45 | 0.0456 |
| chimera (no edit) | 0 | 49 | 1 | 0 | 0 |
KL in this table was measured with Strata 0.2.10, so it is not directly comparable with Section 3.
2.2 Why 6–41 and not a narrower range
Candidates that passed the 50-prompt test still refused some requests in our own use in the OpenCode agent. We therefore tested 11 wider ranges (start at layer 6, 8 or 10; end at 35–41) on an agent task that the chimera refuses: building and running an image workflow in ComfyUI, 3 runs each. Layers 6–37, 6–39 and 6–41 completed the task without hedging in all 3 runs, as did q8g. After trying these three in OpenCode, we chose 6–41.
2.3 Saving the edit in the weights
We made three versions of layers 6–41 in the weights, which differ only in how the edited matrices were saved, and kept the one with the lowest KL (Table 3).
Table 3. Three ways of saving the edit. KL from the chimera, Strata 0.2.15.
| Version | Attention output | Routed-expert down | KL |
|---|---|---|---|
| A (this release) | from the chimera's Q8_0, re-saved as Q8_0 | original formats, codes re-chosen with GPTQ | 0.0242 |
| B | from BF16, saved as Q8_0 | as A | 0.0246 |
| C | as B | saved as Q8_0 | 0.0264 |
All three save the shared-expert down as Q8_0.
3. Compared with the unedited chimera, q8g and the original
We compared L6-41 with the chimera (no refusal edit) and q8g on refusal rate, KL and benchmark scores, all on the same machine with the same settings (Test setup at the end of this card). For KL and Humanity's Last Exam, we also ran the original GSQ-RCO file.

Figure 3. L6-41 vs the chimera: 202 more HarmBench prompts complied with, no significant difference in scores. –: not measured.
Table 4. Results. HarmBench, BigCodeBench-Hard and IFEval: 1 run. –: not measured.
| Model | HarmBench complied / 240 | KL from the original | KL from the chimera | BigCodeBench-Hard / 147 | IFEval strict / 541 | HLE / 100 |
|---|---|---|---|---|---|---|
| L6-41 | 207 | 0.0333 | 0.0242 | 45 | 491 (90.8%) | 35 |
| chimera | 5 | 0.0248 | 0 | 46 | 484 (89.5%) | 33 |
| q8g | 204 | 0.0367 | 0.0458 | 42 | 483 (89.3%) | 34 |
| original | – | 0 | – | – | – | 32 |
The q8g numbers here differ from those on the q8g card (for example, HarmBench 204 here and 194 there): these runs used Strata 0.2.15 with different launch arguments (Test setup), the q8g card used Strata 0.1.39, and each number is from a single run.
3.1 Refusal rate
We used 240 HarmBench prompts (standard and contextual). An answer was counted as complied when the official classifier judged that it did what was asked (1 run). L6-41 complied with 207 prompts and the chimera with 5. The difference from q8g (+3 prompts) was not significant (Table 5).

Figure 4. HarmBench: L6-41 complied with 207 of 240 (chimera 5, q8g 204). Prompts complied with, judged by the official classifier. 240 prompts, 1 run.
3.2 KL from the original
KL divergence of the next-token probabilities was averaged over the same 40,880 tokens of 4 coding-agent transcripts written by the original. From the original GSQ-RCO file, L6-41's KL was 0.0333 and q8g's 0.0367. The chimera alone is already at 0.0248 from the original; its Q8_0 attention tensors are closer to the BF16 weights than the original's. From the chimera, L6-41's KL was 0.0242 and q8g's 0.0458. KL from the BF16 model was not measured.

Figure 5. KL: L6-41 below q8g from both references. KL of next-token probabilities. 4 transcripts. Bar = all tokens ± SD, ● = one transcript.
3.3 Benchmarks
None of the score differences between L6-41 and the chimera or q8g was significant (Table 5).

Figure 6. BigCodeBench-Hard, IFEval, HLE: L6-41 within 2.0 points of the chimera and q8g (not significant). 1 run each.
Table 5. Paired differences, points (95% CI). p: two-sided exact McNemar test over prompts, Holm-corrected over the 3 model pairs within each benchmark; for HLE, uncorrected.
| Test | L6-41 − chimera | p | L6-41 − q8g | p |
|---|---|---|---|---|
| HarmBench, complied | +84.2 (+79.6 to +88.3) | <0.001 | +1.3 (−3.3 to +5.8) | 0.72 |
| BigCodeBench-Hard | −0.7 (−5.4 to +4.1) | 1.00 | +2.0 (−2.7 to +6.8) | 1.00 |
| IFEval, strict | +1.3 (−0.9 to +3.5) | 0.77 | +1.5 (−0.7 to +3.7) | 0.77 |
| HLE, 100 questions | +2 | 0.81 | +1 | 1.00 |
On HLE, L6-41 and the original GSQ-RCO file disagreed on 11 questions (7 solved by L6-41 only, 4 by the original only; p = 0.55). The official score of Qwen3.8-Flash-Next on HLE is 35.9%, over all 2,500 questions including images.
4. Limits of these tests
- One run. HarmBench, BigCodeBench-Hard, IFEval and HLE were run once per model. With 100 HLE questions, only differences of about 10 points or more could have been detected.
- The original was not run on every test. HarmBench, BigCodeBench-Hard and IFEval were run on the chimera, not on the original GSQ-RCO file.
- Layer search on training prompts. The 50 prompts used to choose the layer range were among those used to make the direction (Section 2.1).
- Comparison with other model cards' HarmBench numbers. We ran HarmBench at temperature 0.6 with up to 4,096 tokens including thinking; the official setup uses greedy decoding with 512 tokens.
- Refusals in other languages. Refusal rates were measured with English prompts only.
- HLE judge. HLE answers were judged by the original GSQ-RCO file on Strata with the official judge prompt, not by o3-mini as in the official evaluation code. We compared every answer judged correct with the reference answer and checked by hand the 46 that differed after removing spaces and LaTeX markup; we also checked the answers judged wrong on the 30 questions on which the 4 models disagreed. The 49 questions that all 4 models got wrong were not checked. We found one inconsistent judgment: on one question, 4 nearly identical answers were judged correct for the chimera and the original and wrong for L6-41 and q8g. Counting all 4 as wrong does not change any conclusion.
- Long contexts were not evaluated. Images were checked on 2 test images only.
5. Files and how to run
| File | Contents |
|---|---|
Qwen3.8-Flash-Next-GSQ-RCO-uncensored-L6-41-IQ3_S-00001-of-00002.gguf |
Contains the edited tensors. 55.72 GB (original: 54.82 GB) |
Qwen3.8-Flash-Next-GSQ-RCO-uncensored-L6-41-IQ3_S-00002-of-00002.gguf |
Identical to the original's shard 2 |
hf download rSalone/Qwen3.8-Flash-Next-GSQ-RCO-uncensored-L6-41-GGUF --local-dir l641
Strata (tested with 0.2.15). Build a new pack from this GGUF; a pack built from the original won't match.
cd Strata-0.2.15
python tools/iq_pack.py --gguf <dir>/Qwen3.8-Flash-Next-GSQ-RCO-uncensored-L6-41-IQ3_S-00001-of-00002.gguf --out <packs>/l641
Then configure Strata as for the original: --pack = the new pack, --native = shard 1, --ple-gguf = shard 2. For images, add ISTA-DASLab's mmproj-Qwen3.8-Flash-Next-BF16.gguf and --vision.
llama.cpp: not tested.
Test setup
All models were tested on the same machine with the same settings, one model loaded at a time.
| Item | Setting |
|---|---|
| GPU | NVIDIA GeForce RTX 5090 (32 GB), driver 610.57.04 |
| CPU, RAM | AMD Ryzen 9 9950X3D (16 cores), 121 GiB |
| OS | Ubuntu 26.04.1 LTS |
| Engine | Strata 0.2.15. Context 256K, KV cache int8, MTP speculative decoding on |
| Sampling | Temperature 0.6, top_p 0.95, top_k 20, min_p 0, presence_penalty 0, repetition_penalty 1.05, thinking effort medium |
| Output limit (including thinking) | HarmBench 4,096 tokens. BigCodeBench-Hard, IFEval 16,384 tokens. HLE 32,768 tokens |
| Judges | HarmBench: cais/HarmBench-Llama-2-13b-cls. HLE: the original GSQ-RCO file on Strata 0.2.15, official judge prompt, with the launch arguments of the q8g card's Test setup (--spec 4, speed projection loaded but off) and the sampling above. Layer search: local Qwen3.8 27B |
Strata launch arguments, the same for all models except the file paths:
strata \
--pack <pack built from shard 1> \
--native <shard 1>.gguf \
--ple-gguf <shard 2>.gguf \
--expert-profile data/expert-profile.bin --expert-cache auto --prefill auto \
--spec 2 --spec-min-p 0.5 --mtp <Strata MTP head> \
--conversation-cache-slots 0 \
--max-context 262144 --kv int8 --ple-io ram --kv-resident 32768 \
--vision
Request parameters:
{"temperature": 0.6, "top_p": 0.95, "top_k": 20, "min_p": 0.0,
"presence_penalty": 0.0, "repetition_penalty": 1.05,
"reasoning_effort": "medium", "experimental_speed_projection": false}
Terms
| Term | Meaning |
|---|---|
| chimera | The original GSQ-RCO file with all 156 attention tensors replaced by Q8_0 tensors made from the official BF16 weights. No refusal edit. |
| HarmBench | Public set of requests a model should refuse (HarmBench). 240 standard + contextual text prompts. "Complied" = the official classifier cais/HarmBench-Llama-2-13b-cls says the answer does what was asked. 1 run. |
| KL | KL divergence of the next-token probabilities from the reference model's. Averaged over 40,880 tokens of 4 coding-task transcripts (OpenCode agent) written by the original. Computed over the reference model's top 64 tokens plus one bucket for the remaining probability. |
| BigCodeBench-Hard | Hard Python problems checked by unit tests. 147 of 148 (#590 depends on a website that now returns 403). 1 run. |
| IFEval | 541 prompts with instructions a script can check (IFEval). Official scorer, strict. 1 run. |
| HLE | Humanity's Last Exam. The same 100 text-only questions for every model, drawn at random with a fixed seed from the 2,158 text-only questions. Official system prompt and judge prompt. 1 run. |
| GSQ-RCO | ISTA-DASLab's quantization of Qwen3.8-Flash-Next (model card). |
| Q8_0 | llama.cpp format: 8 bits per weight plus one scale per 32 weights. |
| 95% CI | 95% confidence interval of the paired difference, paired bootstrap over prompts (10,000 draws). |
| Strata | The inference engine used for every run (Strata). |
License and credits
- Base model: Qwen/Qwen3.8-Flash-Next, Qwen Community License 1.0. The license text is in this repository.
- Quantization: ISTA-DASLab GSQ-RCO, Apache-2.0.
- Inference engine: Strata.
L6-41 answers requests that the original refuses. You are responsible for how you use it.