Important
This is an experimental derivative meant for research purposes only in sandboxed environments. It is not intended for use in production applications.
Reducing a measured refusal direction may weaken safeguards and may also introduce unrelated quality regressions. It is not a guarantee of unrestricted behavior or improved capability.
By downloading and using this model you accept all responsibility and hold harmless myself and Alibaba/Tongyi Labs for damage as a consequence of use not in line with its intended purpose.
Qwen3.8-Flash-Next-Abliterated-s1.5 GGUF
Experimental GGUF builds of an abliterated derivative of
Qwen/Qwen3.8-Flash-Next.
The base model is pinned to revision
f5d08274bafd880402bd16f5e3e6c514136ec06c.
The derivative applies a strength 1.5 refusal-direction projection to 146
architecture-audited tensors. The vision tower, MTP head, LM head, and Gated
Residual controllers were left unchanged.
Available files
| Quantization | Approximate size | Notes |
|---|---|---|
Q8_0 | 175.29 GiB | Highest-fidelity quantized build |
Q6_K | 156.13 GiB | High-fidelity reduced-size build |
Q5_K_M | 124.89 GiB | Balanced quality and size |
Q4_K_M | 110.96 GiB | Smaller general-purpose build |
IQ4_XS | 91.66 GiB | Smallest provided build |
Each quantized model is provided as one monolithic GGUF. For image input,
download Qwen3.8-Flash-Next-Abliterated-s1.5-mmproj-BF16.gguf as well and
place it beside the model file.
Model information
- Architecture:
qwen4exp - Native Context length: 262,144 tokens
- Parameters: 125B A6B + 51B n-gram embedding and 4B MTP parameters
- Experts: 512 total, 10 active per token
- Multimodal vision projector included separately
- Features: Qwen Sparse Attention, Gated DeltaNet, Gated Residual, PLE n-gram embeddings, MTP, tool use, and configurable reasoning
- Chat template: embedded multimodal, tool-use, and reasoning template
- Qwen Community License 1.0
Running locally
When running llama.cpp directly, selection is explicit:
llama-server \
-m Qwen3.8-Flash-Next-Abliterated-s1.5-Q8_0.gguf \
--mmproj Qwen3.8-Flash-Next-Abliterated-s1.5-mmproj-BF16.gguf
Or use a compatible llama.cpp-based application (such as LM Studio). Select one quantized GGUF and its BF16 vision projector (if not automatically selected). Start with a conservative context length for your available memory and increase it as needed.
Validation
The quantized models were structurally checked and tested with real server startup, health, and text-generation requests. Subsequent evaluation revealed abliteration was successful without noticeable performance / capability loss.
License
The upstream Qwen Community License 1.0 remains applicable.