license: other
license_name: qwen-community-1.0
license_link: LICENSE
base_model:

• unsloth/Qwen3.8-Flash-Next-GGUF
• orcarouter/Qwen3.8-Flash-Next-Uncensored
  base_model_relation: merge
  tags:
• gguf
• qwen
• qwen3.8
• flash-next
• abliterated
• uncensored

Qwen3.8-Flash-Next-Uncensored-GGUF

This is a community GGUF derivative built from Unsloth Dynamic v3.1 UD-IQ4_XS, incorporating abliterated main-model weights from OrcaRouter's Qwen3.8-Flash-Next-Uncensored.

Only UD-IQ4_XS is provided. The built-in PLE table is included; an external PLE file or runtime LoRA adapter is not required.

Sources and acknowledgements

• Qwen/Qwen3.8-Flash-Next — the original foundation model.
• unsloth/Qwen3.8-Flash-Next-GGUF — the Dynamic v3.1 quantized base.
• orcarouter/Qwen3.8-Flash-Next-Uncensored — the abliterated source weights.

Many thanks to the Qwen, Unsloth, and OrcaRouter teams for making these projects available. This release builds on their work.

This is an independent community derivative, not an official release by those teams.

How this model was made

146 main-model tensors from OrcaRouter's BF16 checkpoint were quantized to the corresponding tensor formats in the Unsloth v3.1 GGUF and substituted in place. Tensor shapes, quantization types, shard layout, and all bytes outside the replacement ranges were preserved.

DeltaNet output-projection columns were converted to GGUF order. The available Unsloth importance matrix was used for applicable tensor types. Q2_0 required a local quantization adapter; the resulting tensor format is unchanged. This does not reproduce Unsloth's complete original quantization recipe.

No additional training or independent refusal-direction search was performed for this release. The transplant covers the main model; it does not transplant OrcaRouter's MTP modifications or reproduce the complete Orca checkpoint.

The inputs were pinned to these revisions:

Files and use

Download all four files from UD-IQ4_XS/ and keep them in the same directory. Load the first shard:

UD-IQ4_XS/Qwen3.8-Flash-Next-UD-IQ4_XS-00001-of-00004.gguf

The four shards total 93,672,359,738 bytes (about 93.7 GB / 87.24 GiB). SHA256SUMS lists their checksums. To verify a local download from the repository root:

sha256sum -c SHA256SUMS

UD-IQ4_XS is a mixed-quantization format. The inference engine must support this model architecture and all tensor formats used by Unsloth Dynamic v3.1. Local comparisons were run with Gufo and llama.cpp; this is not a claim of compatibility with every release or frontend.

No external PLE, MTP sidecar, or vision projector is bundled with this repository. Additional files and engine support may be needed for optional MTP or vision features.

Evaluation scope

Local checks covered file integrity, quantization consistency, short functional tasks, and a limited set of benign prompts that can trigger over-refusal. Factual errors and output-format failures were also observed. These checks do not establish unchanged general capability or the absence of all refusals. Upstream benchmark scores are not evaluation results for this derivative.

License

The model is distributed under the Qwen Community License 1.0. See LICENSE for the full terms, including attribution and commercial-use conditions.

The LICENSE files in both the pinned Qwen and OrcaRouter source revisions contain this license. The OrcaRouter model card labels its license as Apache-2.0; this repository follows the actual LICENSE files rather than that metadata label.

日本語概要

Unsloth Dynamic v3.1のUD-IQ4_XSを土台に、OrcaRouterのQwen3.8-Flash-Next-Uncensoredからmainモデル146テンソルの加工済み重みを移植した、Uncensored / abliterated派生GGUFモデルです。UD-IQ4_XSのみを公開し、内蔵PLEを含みます。

基盤モデルを公開したQwen、量子化モデルを提供したUnsloth、abliterationの成果を公開したOrcaRouterの皆さまに感謝します。