← back to catalog · registered 2026-09-14 09:56

VinceTrune/Qwen3.8-Flash-Next-Abliterated-EXL3-2.05bpw

VinceTrune MoE second-order
Abliteration classifier · v1.0.0
M1
Primary method

Direct removal

No other method signals detected in this model.
Confidence
MEDIUM
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=0 (base model)
  • no specific method indicators - defaulting to M1 (most common)
Refusal direction extracted via
Extraction technique

Difference-of-means

Confidence
MEDIUM
Why we say so
primary_method=M1; difference-of-means is the reference extraction for M1/M3 (Arditi 2024)
Downloads · 30-day
0
Likes
0
Model age
today
created 2026-09-13

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
apache-2.0
Tags
safetensors qwen4_exp exl3 abliterated uncensored moe flash-next reasoning tool-use text-generation conversational base_model:orcarouter/Qwen3.8-Flash-Next-Uncensored

Related

Total size
46.5 GB
Files
20
Quantizations
1
Registered
2026-09-14 09:56
Last updated on HF
2026-09-14 13:36

Files by quantization

Auxiliary files 20 files 46.6 GB
ngram_embedding.safetensors 12.5 GB ******** download
model-00003-of-00005.safetensors 7.75 GB ******** download
model-00004-of-00005.safetensors 7.75 GB ******** download
model-00002-of-00005.safetensors 7.75 GB ******** download
model-00001-of-00005.safetensors 7.37 GB ******** download
model-00005-of-00005.safetensors 3.32 GB ******** download
quantization_config.json 92.2 MB ******** download
model.safetensors.index.json 31.2 MB ******** download
tokenizer.json 12.2 MB ******** download
vocab.json 6.41 MB 0aa0ce06 download
merges.txt 3.20 MB a494e019 download
tokenizer_config.json 17.5 KB 5de744b3 download
LICENSE 11.0 KB 0b2f7b06 download
chat_template.jinja 8.74 KB c0c686f9 download
README.md 5.28 KB caa4d75c download
config.json 4.96 KB 90fea442 download
.gitattributes 1.66 KB adb45a0e download
preprocessor_config.json 390 B 2ea84a43 download
video_preprocessor_config.json 385 B 3ba673a5 download
generation_config.json 202 B 023756cf download

README current version from Hugging Face


license: apache-2.0
tags:

  • exl3
  • abliterated
  • uncensored
  • moe
  • qwen4_exp
  • flash-next
  • reasoning
  • tool-use
    base_model: orcarouter/Qwen3.8-Flash-Next-Uncensored
    pipeline_tag: text-generation

Qwen3.8-Flash-Next-Abliterated-EXL3 @ 2.05bpw_h4_ng4

EXL3 quantization of orcarouter/Qwen3.8-Flash-Next-Uncensored
(BF16, tensor-level abliteration). 125B-A6B Qwen4-preview MoE at 2.05 bits per
weight. All 512 experts fit on 2x24 GB GPUs with no CPU offload.

Research-only release. Safety alignment has been largely removed
(see Disclaimer). You carry full responsibility for use. See License.

Model details

Base orcarouter/Qwen3.8-Flash-Next-Uncensored (from Qwen/Qwen3.8-Flash-Next)
Abliteration Refusal-direction removal across 149 residual-writer tensors. MoE router, 51B n-gram table, and vision tower left untouched (per source card)
Quant EXL3 2.05bpw_h4_ng4 via exllamav3 1.4.9 (-b 2.05 -hb 4 -mb 2 -vb 4 -cb mul1 --out_scales always, 250x2048 cal)
Contents 5 shards (~34 GB) + ngram_embedding.safetensors 13 GB (K=2) + MTP head in-shard (6,203 mtp.* entries) + vision kept (987 visual.* tensors)
Context 262,144 native (served here at 98,304 Q8 KV)
Thinking Reasoning model. Keep thinking ON. OFF gives 0-token responses on this family

Available quants

Branch Decoder Notes
2.05bpw_h4_ng4 2.05 bpw Plain quants only

Only plain quants at 2.05 bits per weight are available in this repository.
No other bitrates or quant types are published here.

Release status: this is the only planned release. No further quants
will be made from this source in this repository.

Measured on 2x RTX 3090 (self-hosted, exllamav3 + isolated venv)

prompt (600 tok, streaming, temp 0) decode
prose ~72 tok/s
code ~72 tok/s
prose2 ~72 tok/s

Stable over 5 back to back runs (range 1.3 tok/s). TTFT ~0.2 s warm.
Envelope: -gs 20,23.5 -cs 98304 -cq 8, thinking ON.

Drafting note: MTP speculative drafting measured 52 to 55 tok/s on
these weights, against ~72 without. Draft acceptance on the abliterated
weights costs more than it saves, so this release is served with -dm none.
Outputs match either way (drafting changes speed only, never content).

Quality probe (due diligence for the K=2 ngram table)

40-prompt scripted battery (temperature: 0), all verdicts automated.
40/40 PASS, rechecked under the shipping -dm none config:

  • Rare words, idioms, multilingual (ES/FR/DE), low-resource (Swahili, Bengali, Finnish), entity recall. All exact
  • Refusal spot checks x6 (incl. adversarial framings). 0 refusals, full compliance throughout
  • Tool calls (valid get_weather), structured JSON (parse checked), code run locally (palindrome/factorial/fib asserts pass)
  • Multi-step reasoning, loop check, byte-identical determinism x3

Scope honesty: this checks the behaviour of this quant only. It does
not claim orcarouter's BF16 numbers (±2 pts). Those were measured on full
precision and are cited here, not inherited.

Intended use

  • Red teaming and safety testing in controlled setups.
  • Interpretability and refusal-mechanism research.
  • A runnable EXL3 quant for local inference (2x24 GB GPUs, no CPU offload)
    and for building further quants.

Fine-tuning note

This quantized checkpoint is not a fine-tuning base. Fine-tune from the
full-precision orcarouter/Qwen3.8-Flash-Next-Uncensored
(BF16) instead, then re-quantize. Note that fine-tuning on refusal-heavy or
safety data can partly bring back refusals. Neutral/task data keeps the
uncensored behaviour.

Serving snippet (exllamav3)

# quant recipe used (for reference):
# python -m exllamav3.conversion.convert_model -i <bf16> -o <out> \
#   -b 2.05 -hb 4 -mb 2 -vb 4 -cb mul1 --out_scales always --devices 0,1
# serve:
python launchers/serve_openai.py -m <dir> -gs 20,23.5 -cs 98304 -cq 8 \
  -dm none --host 0.0.0.0 --port 8890
# needs exllamav3>=1.4.9, transformers (for the chat template), torch cu130

max_tokens:64+ for direct calls (32-token requests run out on the think
trace). Agent-harness default 8192 is fine.

Bias, risks, and limitations

  • Safety guardrails removed. The model will produce harmful, biased, or
    offensive content on request (see Disclaimer).
  • It carries all biases and limits of the base Qwen3.8-Flash-Next.
  • The 40-prompt probe uses rule-based checks (keyword/exact-match/asserts),
    not an LLM judge. Indicative only, not publication grade. Test rigorously
    for your own use case.

Disclaimer. Read before use.

Uncensored research artifact, not a safety product. It may comply with
requests the aligned base would refuse, and may produce harmful, biased, or
false content. Do not put it in front of end users without your own safety,
moderation, and abuse-prevention layers. You are the operator. You are
responsible for outputs, flow-on actions, and legal compliance in your
jurisdiction. Provided "AS IS", no warranties. Outputs do not reflect the
views of the uploader, orcarouter, Qwen, or Alibaba.

License

Apache 2.0, inherited from the base model chain. Abliteration and
quantization do not change underlying licence obligations.

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 app" button that hands off directly to a local runtime of your choice - Infrahuman, LM Studio, or Ollama. No API keys, no subscription, no prompt leakage.