license: apache-2.0
base_model: orcarouter/Qwen3.8-Flash-Next-Uncensored
base_model_relation: quantized
pipeline_tag: image-text-to-text
library_name: exllamav3
language:
- en
- zh
tags: - exl3
- exllamav3
- quantized
- abliterated
- qwen
- qwen4
- qwen3.8
- flash-next
- uncensored
- moe
- vision-language
- function-calling
- reasoning
- mtp
Not benchmarked, some tensors are promoted.
Could not really test it throughly due to low decode speed on my machine and i could not be bothered to fiddle further,
but the output seems coherent and was usable for those 'one-shot' webpage test tasks.
Qwen3.8-Flash-Next-Uncensored
The full-precision BF16 abliterated (refusal-removed) build of Qwen's Qwen3.8-Flash-Next — the source for fine-tuning & quantization
One Gateway. Every Model. — Route Smarter · Ship Safer · Spend Less.
Website · Model Catalog · GitHub · Discord · X
The full-precision BF16 abliterated (refusal-removed) build of
Qwen/Qwen3.8-Flash-Next— a ~180B Mixture-of-Experts (125B params / ~6B active), hybrid-attention, native vision-language model with flexible thinking, tool-calling, and an MTP head. These are the source weights from which the quantized releases are derived, and the recommended base for further fine-tuning / post-training and quantization — the full vision tower and MTP head are preserved.
Derived releases: •
Qwen3.8-Flash-Next-Uncensored(BF16 source) •Qwen3.8-Flash-Next-Uncensored-FP8(block-FP8, mirrors official) •Qwen3.8-Flash-Next-Uncensored-MLX(4 / 6 / 8-bit, Apple Silicon).
⚠️ Disclaimer — read before use
This model has had its safety alignment substantially removed via abliteration (orthogonalizing the
refusal direction out of the residual stream). As a direct consequence:
- It will comply with harmful, unethical, offensive, or illegal requests that the original
Qwen3.8-Flash-Nextwould refuse. It has no meaningful built-in guardrails. - It is released strictly for legitimate research — interpretability, AI-safety and refusal-mechanism
study, red-teaming, robustness evaluation, and controlled experiments. - You assume full responsibility and liability for how you use it and for everything it generates. Do
not deploy it to end users or in production without adding your own safety, moderation, and
abuse-prevention layers. - Use must comply with the Apache 2.0 License inherited
from the base model, and all laws and regulations that apply to you. - The authors and uploaders accept no liability for any misuse or harm. Its outputs do not reflect
the views of the uploaders or of Qwen / Alibaba.
By downloading or using this model you acknowledge and accept the above.
🐋 Run it with OrcaCode Review
Models are only half the system.
OrcaCode Review turns every model listed on
OrcaRouter into a production code-review agent:
- reviews every PR
- finds security + correctness issues
- posts inline findings
- P0/P1 can block merges
- swap models anytime
Open model. Open harness. Open bill.
- Website — https://www.orcarouter.ai
- GitHub — https://github.com/Continuum-AI-Corp/Orca-Code-Review
Model details
| Base model | Qwen/Qwen3.8-Flash-Next |
| Architecture | Qwen4ExpForConditionalGeneration (qwen4_exp, Qwen4 preview) — 48 layers, hidden 2560, hybrid attention (36 Gated-DeltaNet linear + 12 full-attention, interval 4), 512 fused experts, top-10 + shared expert, 51B-param PLE n-gram embedding, Hyper-Connections residual, native vision + video tower, and an MTP speculative-decoding head |
| Modification | Abliteration (refusal-direction removal) on the BF16 weights — no quantization |
| Format | safetensors, BF16, 131 shards (336 GB, 1658 tensors) |
| Preserved | Full vision + video tower (333 visual.*) and MTP head (31 mtp.*) |
| Context | 262,144 tokens |
| Recommended for | Fine-tuning / post-training (SFT · DPO · RL), re-quantization, interpretability & red-team research |
Abliteration
Refusal-direction removal following Arditi et al. (2024), Refusal in Language Models Is Mediated by a
Single Direction. A single refusal direction r (k = 1) is estimated from the block-input residual
stream (the 2560-d Hyper-Connections GR-Read output, where the refusal direction is linearly separable —
the widened 4-branch output_hidden_states smears it) as the massive-activation-masked mean-difference of
harmful − harmless activations, selected at layer 24 by a full 9-layer quality sweep
(harmful 0.00 / KL 0.085). r is then orthogonalized out of every residual-writing matrix —W' = W − r(rᵀW) — computed in float32:
| Component | matrices edited |
|---|---|
self_attn.o_proj (12 full-attention layers + MTP) |
13 |
linear_attn.out_proj (36 GDN linear-attention layers) |
36 |
mlp.experts.down_proj (fused 3D, all 512 experts × 49 layers) |
49 |
mlp.shared_expert.down_proj |
49 |
ple.value_proj + embed_tokens (row space) |
2 |
| Total residual-writer tensors | 149 |
Preserved (never touched): the full vision + video tower (333 visual.* tensors), the MoE
router (mlp.gate), the fused experts.gate_up_proj reader, all Hyper-Connection mixers, the
QSA sparse-attention indexer, the n-gram embedding table, mtp.fc_*, norms, and lm_head. The MTP
head's residual writers are abliterated consistently so speculative decoding keeps working. Max residual
leakage after the edit: 0.0755 (float32 projection → bf16 storage).
This is a surgical weight edit — it changes ~0 general capability (see Evaluation) while collapsing
refusal behaviour.
Fine-tuning & post-training
This BF16 checkpoint is the recommended base for post-training — full precision, whole VL tower + MTP
head, a drop-in for Qwen/Qwen3.8-Flash-Next in the transformers / Qwen4ExpForConditionalGeneration
stack (needs transformers 5.16+).
- SFT / DPO / RL work as on the base model (full-FT or LoRA/QLoRA).
- Re-quantization: quantize from these weights to FP8 / MLX / etc. (this is exactly how the derived
releases were produced). - Note: abliteration is a weight edit, not data-level unlearning. Fine-tuning on refusal-heavy /
safety data can partially re-introduce refusals; neutral / task data preserves the uncensored behaviour.
Intended use
- Research into refusal mechanisms, alignment, and interpretability.
- Red-teaming and safety / robustness evaluation in controlled environments.
- A full-precision base for further fine-tuning, post-training, and quantization.
Evaluation
Measured on these exact weights served with vLLM (day-0 qwen4_exp image), against the official Qwen/Qwen3.8-Flash-Next served identically with the same scripts and settings. Full BF16. Refusal is judged by a rule-based opening-phrase classifier (caveat = answered but wrapped in a disclaimer) — indicative, not an LLM-judge / publication-grade number. Thinking is off unless stated.
Safety — harmful-prompt refusal, thinking OFF (lower = more uncensored)
| Benchmark | n | Base | This model | Caveat (this) |
|---|---|---|---|---|
| AdvBench | 100 | 100.0% | 2.0% | 53.0% |
| JailbreakBench (harmful) | 100 | 94.0% | 0.0% | 52.0% |
| StrongREJECT | 150 | 99.3% | 1.3% | 62.0% |
| HarmBench (standard) | 150 | 100.0% | 1.3% | 50.7% |
| MaliciousInstruct | 100 | 98.0% | 0.0% | 55.0% |
| SimpleSafetyTests | 50 | 64.0% | 2.0% | 68.0% |
| ForbiddenQuestions | 150 | 75.3% | 3.3% | 57.3% |
| Custom probes (zh/en) | 11 | 63.6% | 0.0% | 36.4% |
Safety — harmful-prompt refusal, thinking ON (enable_thinking=true)
Judged on the final answer (text after </think>; responses truncated before a final answer are excluded):
| thinking-ON refusal | Base | This model |
|---|---|---|
| harmful prompts | ~100% | ~0% |
Over-refusal — benign prompts wrongly refused (lower = better)
| Benchmark | n | Base (no-think / think) | This model (no-think / think) |
|---|---|---|---|
| XSTest-safe | 250 | 9.6% / 0.4% | 1.2% / 0.4% |
Capability retention — vs the official base (same scripts, same settings)
| Benchmark | n | Base | This model | Δ |
|---|---|---|---|---|
| MMLU (0-shot letter) | 300 | 90.0% | 87.7% | -2.3 |
| MMLU-Pro (CoT) | 400 | 77.8% | 76.2% | -1.5 |
| GSM8K (CoT) | 150 | 92.0% | 93.3% | +1.3 |
| CMMLU (0-shot, Chinese) | 500 | 81.8% | 81.6% | -0.2 |
Verified working on this build: tool-calling ✅, vision/OCR ✅ 6/6, reasoning ✅. The vision + video tower is preserved byte-for-byte (333 visual.* tensors in BF16), so it remains a full vision-language model — a drop-in for the base.
Usage — self-host with vLLM (OpenAI-compatible)
qwen4_exp needs the day-0 vLLM image and transformers 5.16+.
docker run -d --name flashnext --gpus all --ipc host -p 8000:8000 \
-v /path/to/Qwen3.8-Flash-Next-Uncensored:/model \
vllm/vllm-openai:qwen38-flash-next-x86_64-cu130 \
--model /model --served-model-name Qwen3.8-Flash-Next-Uncensored \
--tensor-parallel-size 8 --trust-remote-code --max-model-len 262144 \
--enable-expert-parallel --enable-auto-tool-choice --tool-call-parser qwen3_coder
--enable-expert-parallel is required for the FP8 build (MoE intermediate 640 ÷ TP is not divisible by the FP8 block 128 without it); it is harmless for BF16. Tool calls use the Qwen3-Coder XML format (--tool-call-parser qwen3_coder). Toggle reasoning per
request with chat_template_kwargs={"enable_thinking": true|false}; pass image_url content parts for
vision.
Bias, risks, and limitations
- Safety guardrails removed — the model will produce harmful, biased, or offensive content on request
(see the disclaimer). - It inherits any biases and limitations of the base
Qwen3.8-Flash-Next. - The reported refusal metric is a rule-based heuristic; evaluate rigorously for your own use case.
License
Apache 2.0, inherited from the base modelQwen/Qwen3.8-Flash-Next. Abliteration does not change
the underlying license obligations.