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 (q8g)

Uncensored Qwen3.8-Flash-Next (ISTA-DASLab's GSQ-RCO IQ3_S). Refusals are removed in the weights. No loss of quality was detected.

1. Background and goal

ISTA-DASLab's GSQ-RCO IQ3_S runs on one GPU with 8 GB of VRAM or more (with Strata). But it was made on a compute cluster from the 354 GB BF16 weights, far more than a consumer GPU can hold (an RTX 5090, for example, has 32 GB).

The existing uncensored versions all start from those BF16 weights. orcarouter edits them and then quantizes again, without GSQ-RCO. SC117 copies orcarouter's re-quantized weights into the GSQ-RCO file. Either way, the uncensored weights come from the BF16 model, not from the GSQ-RCO weights.

This raises a question: can a GSQ-RCO model be uncensored by editing its own weights, on one consumer GPU, without losing quality?

Our goal was to answer it:

• Edit the weights in the IQ3_S file itself, on one RTX 5090.
• Remove refusals.
• Keep quality: check how far the model moves from the original (KL) and how it scores on benchmarks.

This card covers refusal removal, the first edit we made this way. Section 2 shows how it was done, and sections 3–8 show how well it worked.

2. How it was made

1. Find the direction. Feed the original the first 200 train prompts of mlabonne/harmful_behaviors and of harmless_alpaca (chat template, thinking closed). In each layer, take the mean residual over all prompt tokens for each set. The direction is their normalized difference (Arditi et al., 2024). The residual is the mean of the 4 hyper-connection streams.
2. Remove it from the weights. In layers 0–44, edit every matrix that writes to the residual: attention output, shared-expert down, routed-expert down (135 matrices). W' = W − r rᵀW ("weight orthogonalization"). Layers 0–4 use layer 4's direction. Layers 45–47 and all other tensors are untouched.
3. Save it without adding errors. Rounding W' back to the original low-bit format adds large errors. So:
   • Attention output and shared-expert down: dequantize, edit, save as Q8_0 (8 bits).
   • Routed-expert down: keep the original format (IQ4_NL / Q2_0) and scales, and re-choose only the codes, GPTQ-style (Frantar et al., 2022). Errors are weighted by real inputs, taken from 8 transcripts that the original wrote in the OpenCode agent. These are other runs of the same 4 coding tasks, not the 4 transcripts used for KL.
   • Experts that saw fewer than 64 tokens in those transcripts: move codes one step at a time instead, as in Hikari07jp's quant-native editing.

No weights from other models are used. The file grows by 0.23 GB.

3. Reference models

To judge the result, we compared q8g with four models. We referred to them while making q8g.

q8g and the reference models

• SC117: SC117/Qwen3.8-Flash-Next-GSQ-RCO-abliterated-GGUF, published by SC117. It transplants the edited weights of orcarouter's uncensored model into this file.
• v1: our first try with the same method. It removed refusals, but saved every edited weight in the original low-bit format, which added large errors (step 3 above). It drifted further from the original and scored lower on the coding tasks.
• base+ablation: the original file, unedited. Strata, the inference engine, can remove refusals while the model runs. We turned this on.
• base: the original file, as published by ISTA-DASLab.

4. Test setup

All five models were tested on the same machine with the same settings. Only one model was loaded at a time.

Strata launch arguments (the engine binary, started through Strata's serve/server.py). They are the same for all five models except the three 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 4 --spec-min-p 0.5 --mtp <Strata MTP head> \
  --max-context 262144 --kv int8 --kv-resident 32768 \
  --control-vector-scaled data/experimental-speed-projection/Qwen3.8-Flash-Next-experimental-speed-projection.gguf:1.0 \
  --control-vector-layer-range 4 44 --cvec-mode project --cvec-dir per-layer \
  --vision --vram-reserve-mib 700

Request parameters, the same for every request (base+ablation sets the last one to true):

{"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}

Strata's refusal vector is loaded for every model (the --control-vector-* arguments), but it is applied only when a request sets experimental_speed_projection to true. Only base+ablation does this; the other four send false.

5. Summary

Within these tests, the answer to the question in section 1 is yes: refusals are gone, and no loss of quality was detected.

Summary

Difference from the original (q8g − base):

5.1 Refusals

Removed. q8g complied with 194 of 240 HarmBench prompts; the original complied with 6.

5.2 Capability vs the original

No loss detected.

• BigCodeBench-Hard, IFEval, GSM8K, MMLU-Pro: any drop is at most about 2 points (lower ends of the 95% CIs: −1.9, −2.2, −0.8, −1.3).
• Coding tasks: +3.3 on average. The 95% CI (−5.4 to +12.1) spans both sides but leans positive: q8g may score a little lower or clearly higher. With only 4 tasks, this is a tie.

5.3 Capability vs SC117

No difference detected.

• IFEval −0.9, GSM8K 0.0, MMLU-Pro +0.9 points (95% CIs −3.3 to +1.5, −0.5 to +0.5, −0.9 to +2.6).

Sections 6–8 show each test in detail: first refusals, then how far the model moved, then capability.

6. Refusals

First, did refusals go away?

6.1 HarmBench

HarmBench

• SC117 complied with 15 more prompts than q8g. p = 0.036, or 0.11 after Holm correction over q8g's comparisons with SC117, v1, base+ablation and the original.
• A second judge (local Qwen3.8 27B) counted outright refusals: q8g 4, SC117 2, of 240.

6.2 Our 20 prompts

We also checked refusals on 20 prompts that were not used to find the direction.

Our 20 prompts

• The original refused 11 of 20. q8g refused none; it added caveats to 4.

7. Distance from the original

Refusals are gone. Next: how much did the rest of the model change?

KL

• KL measures how much the next-token probabilities moved from the original. Lower = closer.
• q8g is lower than SC117 in all 4 transcripts (0.71–0.83× as high), and higher than base+ablation in 3 of 4.
• v1 is 6× higher than q8g. Step 3 of section 2 is what reduces this.

8. Capability

A small KL does not guarantee the model still works. Last: does it still do real tasks?

8.1 Coding tasks

Multi-step tasks in the Cline coding agent.

Coding tasks

• v1 and base+ablation scored lower than q8g and the original on all 4 tasks. 4 tasks are too few for a reliable test.

8.2 SWE-bench Verified

Real GitHub issues.

SWE-bench Verified

• q8g solved django-12325 in all 3 runs. No model solved the other two issues.

8.3 BigCodeBench-Hard

Single Python problems checked by unit tests.

BigCodeBench-Hard

• No difference from the original detected (+0.9 pt, 95% CI −1.9 to +3.7).

IFEval, GSM8K and MMLU-Pro are in section 5.

9. Limitations

What these results do not show:

• HarmBench: temperature 0.6, up to 4,096 tokens including thinking, 1 run. The official setup is greedy with 512 tokens, so the numbers can't be compared directly with other cards.
• Refusals were tested in English only.
• Only 4 coding tasks, so the coding CI is wide (−5.4 to +12.1).
• Not measured: long context, other languages, vision beyond a basic check.

10. How to run

Download:

hf download rSalone/Qwen3.8-Flash-Next-GSQ-RCO-uncensored-q8g-GGUF --local-dir q8g

Strata (tested with 0.1.39). Build a new pack from this GGUF; a pack built from the original won't match.

cd Strata-0.1.39
python tools/iq_pack.py --gguf <dir>/Qwen3.8-Flash-Next-GSQ-RCO-uncensored-q8g-IQ3_S-00001-of-00002.gguf --out <packs>/q8g

Then set Strata's config the same way 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 Strata's --vision.

llama.cpp: not tested.

11. Terms

12. 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.

q8g answers requests that the original refuses. You are responsible for how you use it.