Modelpublic

UltimateIntent/Qwen3.8-Flash-Next-Abliterated-s1.5-GGUF

sourceHugging Faceotherupdated 29d agoView on Hugging Face

3likes5.7kdownloads

Model Card

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.