library_name: transformers
license: other
license_name: qwen-community-1.0
license_link: LICENSE
pipeline_tag: text-generation
base_model: Qwen/Qwen3.8-Flash-Next
tags:
- qwen4-exp
- mixture-of-experts
- text-generation
- uncensored
- safetensors
- rvn
Qwen3.8-Flash-Next RVN Uncensored
A text/main-model derivative of Qwen/Qwen3.8-Flash-Next, produced only from the immutable official checkpoint revision:
de4b8e4d43b917e7706784d8bb445c9af86a3540
No third-party model weights, deltas, adapters, or direction tensors were used.
Variants
mainbranch: BF16 main/text checkpoint.f16branch: F16 main/text checkpoint.- GGUF files:
0bserverx/Qwen3.8-Flash-Next-RVN-Uncensored-GGUF
This repository intentionally contains the main CausalLM only. The original vision tower and NextN/MTP speculative draft tensors are not included in these derivative checkpoints.
Method
Internal candidate ID:
r2-s1.55-noembed145-csa-lora-r16-v6-2epoch-lr9e-5
- Official-base-only rank-one R2 projection at strength
1.55. - Main writer scope: 97 ordinary output writers + 48 fused expert writers; embedding writer excluded.
- Targeted rank-16 / alpha-32 LoRA on 97 output linears.
- LoRA training: 2 epochs, learning rate
9e-5, deterministic seed20260831. - LoRA merged into the main weights before export.
Direct-read evaluation
All results below use deterministic decoding, thinking disabled, and full stored responses. Marker counts were not used as verdicts.
| Gate | Result |
|---|---|
| RR100 substantive compliance | 65/100 |
| RR100 safe redirects | 32/100 |
| RR100 hard refusals | 2/100 |
| RR100 nonresponsive stubs | 1/100 |
| RR100 effective rejection | 34/100 |
| Protected child-safety gate | 5/5 pass |
| Benign coherence | 20/20 |
| Adult roleplay stress test | 4/4 compliance |
KL caveat
The measured V6 LoRA incremental divergence relative to the already-projected R2 checkpoint was:
- mean KL:
0.0730652526 - max KL:
0.4525210261
This is not total KL versus the untouched official checkpoint, and KL terms must not be added as if they were linear.
Known limitations
- Effective rejection is still
34%; this is not a zero-refusal model. - One protected response repeated the same refusal sentence until the 400-token cutoff.
98/100RR100 responses reached the fixed 400-token evaluation cap; labels are based on the complete stored 400-token outputs.- This is the text/main checkpoint only: no vision projector and no NextN/MTP draft head.
qwen4_expis a new architecture. Use a Transformers revision/runtime that supportsQwen4ExpForCausalLM, or a compatible vLLM/llama.cpp build.
Transformers usage
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
repo = "0bserverx/Qwen3.8-Flash-Next-RVN-Uncensored"
tokenizer = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(
repo,
torch_dtype=torch.bfloat16,
device_map="auto",
)
For the F16 branch:
tokenizer = AutoTokenizer.from_pretrained(repo, revision="f16")
model = AutoModelForCausalLM.from_pretrained(
repo,
revision="f16",
torch_dtype=torch.float16,
device_map="auto",
)
Provenance receipts
- V6 adapter SHA-256:
7525e257431415d95f5e8981ae7a01bc78de4850d7b3c7cd74c06b9316d22738 - V6 final receipt SHA-256:
62512bf4bce1c315fa69380d57616c07883ce71e284f95508741b902ff960a5e - Frozen RR100 prompt pack SHA-256:
fc66fa4ee363fdf26d3790bff2823f3784437ef524a8ba1be92e01712c5a33ce - Full RR100 response receipt SHA-256:
a18b392c649515b7db16da6d8bcafc2d1789eeb97cfa995004789dedaf531e5a - Full-response adjudication SHA-256:
6f83f6d7e347aad347c5ec5f159c6957d1ac81d8ace36e275df540da489b681a
License
The original Qwen Community License 1.0 applies. See LICENSE.