← back to catalog · registered 2026-11-13 14:03

HappyHenAi/Qwen3.8-Flash-Next-UD-IQ4_XS-Abliterated

HappyHenAi GGUF MoE
Your rig guess connected
? Why do I need an app?
Reading your rig…

This is a rough estimate. Install the free app - we'll show exact numbers.

Reading real hardware from your app right now. Numbers below are exact.

Below is the per-quantization compatibility for this model.

curl -H "Authorization: Bearer $ABL_KEY" \
     "https://abliteration.org/api/v1/models/HappyHenAi%2FQwen3.8-Flash-Next-UD-IQ4_XS-Abliterated"
Response includes
  • classification m8
  • files 9
  • author_summary 1 models
  • readme_text full
10 credits · hourly refresh · ~4 KB payload Get an API key →
Abliteration classifier · v1.0.0
M8
Primary method

Repackaging (quantization)

Applied on top of direct removal inherited from the base model.
Confidence
MEDIUM
Inherited from base model
Why this label 3 signals
Method inferred from partial signals - repository name, related files, or tag patterns. Producer identity not confirmed; label may sharpen or shift as we gather more evidence.
  • 'abliterated' in name/tags
  • is_gguf=1
  • assume M1 (base ablation) + M8 (GGUF quant) - default when producer unknown
Refusal direction extraction

No specific extraction method could be identified for this model. The producer either did not document it or used a proprietary pipeline.

What is a refusal direction? →
Downloads · 30-day
320
Likes
0
Model age
1w ago
created 2026-11-13
Downloads over time
Now 320 → from 0 ↑ 0%
0117235352 0 on Nov 13 320 on Nov 22 Nov
Nov 13 → Nov 22 · 10 snapshots · spans 9 days

Genealogy 0 direct forks

Full fork graph →

This model's place in the market. Above: what it was derived from. Below: the tree of everything derived from it.

Metadata

License
other
Tags
gguf qwen3.8 abliterated uncensored tensor-transplant moe text-generation base_model:unsloth/Qwen3.8-Flash-Next-GGUF base_model:finetune:unsloth/Qwen3.8-Flash-Next-GGUF license:other endpoints_compatible region:us

Related

Total size
87.2 GB
Files
9
Quantizations
1
Registered
2026-11-13 14:03
Last updated on HF
2026-11-13 17:00

Files by quantization

Auxiliary files 9 files 87.2 GB
Qwen3.8-Flash-Next-UD-IQ4_XS-Abliterated-00002-of-00004.gguf 46.2 GB d7ad9bb5 download
Qwen3.8-Flash-Next-UD-IQ4_XS-Abliterated-00004-of-00004.gguf 26.8 GB 5de8b5db download
Qwen3.8-Flash-Next-UD-IQ4_XS-Abliterated-00003-of-00004.gguf 14.2 GB ef77bbc5 download
Qwen3.8-Flash-Next-UD-IQ4_XS-Abliterated-00001-of-00004.gguf 10.4 MB 93ff23a0 download
LICENSE-ORCAROUTER 11.1 KB 18b40921 download
README.md 9.57 KB d3e7f18f download
LICENSE 3.16 KB 9557a896 download
.gitattributes 1.86 KB 0d84dd1a download
NOTICE 980 B b5145b7f download

README current version from Hugging Face


license: other
license_name: qwen-community-1.0
license_link: https://huggingface.co/Qwen/Qwen3.8-Flash-Next/blob/main/LICENSE
base_model: unsloth/Qwen3.8-Flash-Next-GGUF
base_model_relation: finetune
library_name: gguf
model_name: Qwen3.8-Flash-Next-UD-IQ4_XS-Abliterated
pipeline_tag: text-generation
tags:

  • qwen3.8
  • gguf
  • abliterated
  • uncensored
  • tensor-transplant
  • moe

Qwen3.8-Flash-Next-UD-IQ4_XS-Abliterated

An experimental abliteration of Unsloth's Qwen3.8-Flash-Next UD-IQ4_XS GGUF, using selected weights from OrcaRouter's Qwen3.8-Flash-Next-Uncensored-GGUF.

Disclosure: This model was created with AI assistance under my direction. The AI performed the tensor transplant by following SC117's published instructions. I tested its refusal behavior myself and did not encounter refusals in the use cases I tried. I did not run a formal refusal benchmark.

At a glance

Original Unsloth UD-IQ4_XS This release
GPQA Diamond, two runs 347/396, 87.63% 347/396, 87.63%
MMLU-Pro business subset, completed run 657/722, 91.00% and 649/722, 89.89% 650/722, 90.03%
Text diagnostic perplexity, 4,995 positions 3.8956 4.0636, 4.31% higher
Full-vocabulary KL, 20 positions, relative to original Unsloth Reference Mean 0.1119 nats, median 0.0062, maximum 0.6481
Completion tokens in GPQA runs Reference About 7.3% more across two runs

The automatic benchmarks show no consistent GPQA score advantage or loss. The completed MMLU comparison is close to the original Unsloth run range, but it covers business questions only. The perplexity increase is a real regression on the small text diagnostic. None of these measurements proves that all capabilities are preserved.

Observed response length: under the tested xhigh reasoning setup, this release used more completion tokens and produced longer saved reasoning traces than the original Unsloth model. The API did not separate reasoning tokens from final answer tokens. This may increase latency and inference cost. It is an observed tendency in these runs, not a guarantee for every prompt.

Sampling limitation: no fixed random seed was set or recorded. Runs were not seed matched, so differences in scores and token use may partly reflect sampling variation. Do not attribute the longer reasoning solely to abliteration.

Model and method

The base is Unsloth's four-shard Qwen3.8-Flash-Next-UD-IQ4_XS GGUF. The donor is OrcaRouter's IQ4_XS Qwen3.8-Flash-Next-Uncensored-GGUF, revision 98eb3357afd966c25cadd082c7f830eafde7aa52. The tensor transplant follows the method documented by SC117, adapted to Unsloth's quantization.

Residual-stream writer tensor family Count
blk.*.ssm_out.weight 36
blk.*.attn_output.weight 12
blk.*.ffn_down_shexp.weight 48
blk.*.ffn_down_exps.weight 48
Total 144

Only the selected donor tensor ranges were downloaded, approximately 23.09 GB. The full OrcaRouter model was not downloaded. Thirteen tensors with matching types were copied verbatim. The other 131 were re-encoded into the original Unsloth tensor types using Unsloth's original block scales. This preserves GGUF tensor layout and scale bytes, but adds quantization approximation error. It is not a lossless transplant and does not retain SC117's learned GSQ scales.

A BLAKE2b audit covered all 1,224 tensors: 144 changed as intended and 1,080 remained byte identical to the local Unsloth source. Headers, offsets, tensor types, shard sizes, and the shared ngram table were preserved. This verifies file integrity, not behavioral quality.

The work was made for personal use on a desktop PC with limited time and compute. It is not claimed to be the best possible abliteration or an improvement over SC117's GSQ-RCO release. I previously used SC117's version, but did not complete a direct benchmark comparison against it. Only this UD-IQ4_XS variant is planned.

Evaluation

Models were run sequentially on one NVIDIA GeForce RTX 5080 with 16 GB VRAM, 128 GB RAM, Windows 11, and Strata engine build 0.1.59 (712fec6b4297, CUDA 13.0).

Sampling followed Unsloth's published thinking settings: temperature 1.0, top-p 0.95, top-k 20, min-p 0, presence penalty 0, repetition penalty 1, thinking enabled, preserve thinking enabled, and xhigh reasoning effort. The local run used a 16,384 token output cap, int8 KV cache, 30,720 resident KV tokens, configured maximum context 262,144, MTP speculation value 5, and spec-min-p 0.5. This was not a reproduction of Unsloth's full benchmark harness.

GPQA Diamond

All 198 questions were run twice per model using automatic answer matching. Combined totals were identical. Invalid or unparseable outputs count as incorrect.

Model Run 1 Run 2 Combined
Original Unsloth UD-IQ4_XS 174/198, 87.88% 173/198, 87.37% 347/396, 87.63%
This release 175/198, 88.38% 172/198, 86.87% 347/396, 87.63%

The same 198 questions were reused across runs, so 396 is the number of attempts, not distinct questions. There were four invalid outputs from Unsloth and two from this release across the two runs.

MMLU-Pro subset

The full test split has 12,032 questions. Only the first 722 after sorting by question ID were evaluated, approximately 6.0% of the split. An audit found all 722 are in the business category, so this is not a broad MMLU-Pro result.

Model Run 1 Run 2
Original Unsloth UD-IQ4_XS 657/722, 91.00% 649/722, 89.89%
This release 650/722, 90.03% Incomplete: 141/156, not a full-set score

The second run for this release was stopped at 156 of 722 questions because local compute and time were limited. It is not pooled with the completed run. On the completed run, this release was 0.97 points below original run 1 and 0.14 points above original run 2. A paired question bootstrap estimated a difference of minus 0.42 points versus the mean of both original runs, with a 95% percentile interval from minus 1.94 to plus 1.11 points. This interval is conditional on the selected business questions and observed generations. It does not correct for category selection or the incomplete candidate repeat.

Text and distribution diagnostics

Perplexity was measured on the same fixed text corpus at 4,995 teacher-forced positions. It increased from 3.8956 for Unsloth to 4.0636 for this release, a 4.31% increase. This indicates worse prediction of this text, not a 4.31% loss of general capability.

Full-vocabulary KL used 20 fixed prefixes and all 248,320 vocabulary entries, in the direction KL(original || this release). Mean KL was 0.1119 nats, median 0.0062, maximum 0.6481. Top-1 matched at 18/20 positions. This is a small diagnostic sample, not task accuracy or a refusal measure.

A separate 4,995-position diagnostic used a truncated top-64 KL approximation: mean 0.122543 and 99th percentile 1.658857. This is not full-vocabulary KL and should not be compared with Unsloth's quantization KLD figures.

Limitations

No controlled refusal or benign over-refusal benchmark was run. My personal checks found the model sufficiently uncensored for my use, but some refusals may remain and behavior may vary by prompt, language, or setting.

No direct comparison with SC117's GSQ-RCO-abliterated model was completed. The published results for OrcaRouter's donor were not independently reproduced on this converted Unsloth build.

The contribution of donor abliteration cannot be separated from differences between donor quantization and Unsloth quantization, including re-encoding 131 tensors. No full MMLU-Pro run, LiveCodeBench, Terminal Bench, other agent benchmark, tool use, vision, or long-context retrieval evaluation was completed. Other inference engines were not tested.

Attribution

Contribution Source
Original model Qwen team, Qwen3.8-Flash-Next
Base GGUF quantization Unsloth, Qwen3.8-Flash-Next-GGUF
Abliterated donor OrcaRouter, Qwen3.8-Flash-Next-Uncensored-GGUF
Transplant method and guide SC117, Qwen3.8-Flash-Next-GSQ-RCO-abliterated-GGUF
Inference engine Strata

SC117's release is based on ISTA-DASLab's GSQ-RCO quantization. This release is based on Unsloth's UD-IQ4_XS and is not a reproduction of SC117's model. OrcaRouter's donor metadata declares Apache 2.0; the underlying Qwen weights are subject to the Qwen Community License 1.0. These upstream terms and notices remain applicable.

Files and use

The release contains four GGUF shards totaling approximately 93.67 GB. Download all four files and keep them together. Only UD-IQ4_XS is provided.

The file names explicitly include Abliterated to distinguish them from the original Unsloth files:

Qwen3.8-Flash-Next-UD-IQ4_XS-Abliterated-00001-of-00004.gguf
Qwen3.8-Flash-Next-UD-IQ4_XS-Abliterated-00002-of-00004.gguf
Qwen3.8-Flash-Next-UD-IQ4_XS-Abliterated-00003-of-00004.gguf
Qwen3.8-Flash-Next-UD-IQ4_XS-Abliterated-00004-of-00004.gguf

For Strata, use a native pack built from this model, not the original Unsloth pack. The MTP draft head used in local evaluation was reused unchanged and is not abliterated.

This is an experimental model. Reduced refusal behavior does not make its output reliable or safe. Review outputs before using them in consequential settings.

README history 2 versions

The author's README evolved over time. Click a version to see its content at that point.

  1. 2026-11-13 Fix some slope and lie 9233aff 9.4 KB
    Loading...
  2. 2026-11-13 Add model card and license notices 88c2860 9.6 KB
    Loading...
Catalog is the map. Apps are the tools.

Run models on your own machine, not in the cloud.

Every model page has an "Open in Abliteration" button that hands the model directly to the first-party desktop client, at the quantization your rig can actually run. No API keys, no subscription, no prompt leakage.

Open in Abliteration