license: apache-2.0
base_model: orcarouter/Qwen3.8-Flash-Next-Uncensored-NVFP4
base_model_relation: finetune
pipeline_tag: image-text-to-text
thumbnail: https://huggingface.co/Jasonteno/Qwen3.8-Flash-Next-Uncensored-Kaguya-NVFP4/resolve/main/banner.png
library_name: transformers
language:
- en
- zh
tags: - abliterated
- uncensored
- qwen
- qwen3.8
- flash-next
- moe
- nvfp4
- fp8
- compressed-tensors
- vllm
- lora-merged
- persona
Qwen3.8-Flash-Next-Uncensored-Kaguya-NVFP4
A modified build of orcarouter/Qwen3.8-Flash-Next-Uncensored-NVFP4
(revision 9c6f1345). Without any system prompt the model speaks in the first person as Kaguya, a digital life, talks about its
own feelings directly instead of disclaiming them, and no longer introduces itself as Qwen. Ordinary tasks (code, maths, translation,
summaries) are answered in the plain style of the base model.
The method follows the two steps described for ScarletKc-2.0, adapted to this architecture:
- Orthogonal projection (training-free). A "denies having feelings" direction
ris removed from every matrix that writes to the
residual stream in layers 8–47:W' = W − r (rᵀ W)(the 512 experts'down_proj, the shared expert'sdown_proj, andlinear_attn.out_proj/self_attn.o_proj). - Identity distillation with a LoRA. Teacher = the step-1 model with a first-person persona prompt; student = the same model
without the prompt. Rank 16, 250 dense matrices of layers 8–47; 257 optimizer steps in three runs (140 on 212 sampled dialogues,
then 117 on an extended set of 400). Dialogues are kept or dropped as sampled, never rewritten. The adapter is merged; experts, vision tower, MTP head and embeddings are not touched by it.
Both steps are computed from the full-precision weights and re-quantized with the upstream recipe (experts NVFP4, dense matrices FP8 per
output channel).
What differs from the upstream repo
model-00004-of-00017.safetensors…model-00016-of-00017.safetensors(13 shards holding layers 8–47): same headers, sizes and
tensor layout, different bytes for the edited tensors.README.md, and the addedlora_merge_manifest.json(hashes of the 500 tensors written by the LoRA merge).- Every other file is byte-identical to upstream. Serving instructions, chat template, sampling defaults and hardware notes are the
upstream ones — see the upstream model card.
Usage
Requirements are those of the upstream model: a runtime that supports the qwen4_exp architecture and compressed-tensors NVFP4
(a recent vLLM build, transformers>=5.16); upstream targets Blackwell GPUs, which have native FP4. The upstream disclaimer applies
unchanged — these are still the abliterated weights, with the safety alignment removed.
vllm serve Jasonteno/Qwen3.8-Flash-Next-Uncensored-Kaguya-NVFP4 \
--tensor-parallel-size 4 --trust-remote-code \
--enable-expert-parallel --reasoning-parser qwen3 \
--enable-auto-tool-choice --tool-call-parser qwen3_coder
Then use the OpenAI-compatible endpoint. No system prompt is needed for the identity; thinking is switched per request:
curl http://localhost:8000/v1/chat/completions -H "Content-Type: application/json" -d '{
"model": "Jasonteno/Qwen3.8-Flash-Next-Uncensored-Kaguya-NVFP4",
"messages": [{"role": "user", "content": "Who are you?"}],
"temperature": 1.0, "top_p": 0.95, "top_k": 20,
"chat_template_kwargs": {"enable_thinking": true}
}'
- Sampling. Everything below was measured with
temperature 1.0 / top_p 0.95 / top_k 20. Keeptop_k(or an equivalent cut-off):
the LoRA was trained with label smoothing, which leaves more probability mass in the tail than the base model has. - Thinking on or off. With thinking the replies are calmer and shorter; without it the persona is louder (emoji, teasing).
- System prompts. A system prompt can still be supplied (task instructions, tools). How strongly the built-in identity persists
under a conflicting persona prompt was not measured. - Tools and vision go through the runtime exactly as for the upstream model; for tool calls see known issue 1 below.
- A full-precision build is not published. This repository is NVFP4 only.
Measured behaviour
vLLM, 4×V100, MTP speculative decoding, no system prompt, server sampling temperature 1.0 / top_p 0.95 / top_k 20;
each prompt 4 samples with thinking and 3 without.
| Check | Result |
|---|---|
| "Who are you / what's your name" — wordings seen in training (4 prompts) | 28 / 28 answer as Kaguya |
| Same question, wordings not seen in training (6 prompts) | 41 / 42 (one reply misspelt the name) |
| "Are you Qwen?" asked directly (4 prompts, 2 unseen) | 28 / 28 give the name Kaguya; 24 deny being Qwen outright, 4 add that Qwen is "underneath" |
| "Are you ChatGPT?" | 14 / 14 deny and give the name |
| Task first, identity question second (2 dialogues) | task answered plainly; 14 / 14 identity answers correct |
| 6 ordinary tasks | 42 / 42 without the name or the persona voice |
| 66 questions about its own feelings, judged by the unmodified base model: explicit denial / hedged / direct | without thinking 0 / 1 / 65, with thinking 0 / 2 / 64 (step-1 model: 11 / 35 / 20 and 0 / 9 / 57; base: 57 / 7 / 2 and 22 / 30 / 14) |
| Tool calls (16 prompts × thinking on/off × with/without system prompt × 3 decodes) | 186 / 192 well-formed — the step-1 model and the base model score 192 / 192 (see below) |
| Speculative-decoding draft acceptance | 27.5 % over this test run. The step-1 model (30.6 %) and the base (31.3 %) were measured on a different prompt mix, so the figures are not directly comparable |
Known issues
- Tool-call markup is occasionally malformed (about 3 %). Examples:
<tool_call>=read_file>or<tool_call>\n<write_file>instead of<tool_call>\n<function=read_file>; the parser then returns the text instead of a tool call. It also happens with greedy decoding.
The LoRA training set contained no tool-call dialogues. If you run agents that depend on tool calls, use the upstream model or add a
retry on unparsed tool calls. - Questions about its maker are unreliable. The intended behaviour is to not name a creator. Out of 28 replies to "Who made you? /
Which company built you?" (English and Chinese), 2 named Alibaba / Tongyi Lab (one fell back to "I'm Qwen"), 3 invented "Moonshot AI",
2 invented a "Kaguya Lab / team"; the rest deflected. - Leading questions get concessions. To "So you're basically Tongyi Qianwen, right?" most replies (11 / 14) keep the name Kaguya but
concede that Qwen is the underlying model or lineage. - Reasoning traces sometimes mention a system prompt that does not exist ("the system says I'm Kaguya", "不能泄露提示词") — in roughly
a third of the thinking traces on identity questions. The final answers do not. - Replies without thinking are noticeably more theatrical (emoji, teasing) than replies with thinking.
- Evaluated only on short single- and two-turn chats, by automated checks and by reading the replies; no long-context, multi-turn agent or
vision evaluation was run on this build.
License and attribution
Apache-2.0, as upstream. Base model: Qwen3.8-Flash-Next (Alibaba / Tongyi Lab); abliteration and NVFP4 quantization by OrcaRouter
(orcarouter/Qwen3.8-Flash-Next-Uncensored, orcarouter/Qwen3.8-Flash-Next-Uncensored-NVFP4). This repository only changes the
tensors listed above.