base_model: dealignai/Qwen3.8-Flash-Next-UNCENSORED-FP8 library_name: llama.cpp pipeline_tag: text-generation license: other license_name: qwen-community-license-1.0 license_link: https://huggingface.co/dealignai/Qwen3.8-Flash-Next-UNCENSORED-FP8/blob/main/LICENSE tags:
- gguf
- qwen
- qwen4_exp
- iq2_xxs
- uncensored
- abliterated
Qwen3.8-Flash-Next-UNCENSORED IQ2_XXS GGUF
IQ2_XXS GGUF conversion of
dealignai/Qwen3.8-Flash-Next-UNCENSORED-FP8, intended as a practical build for a system with 64 GB RAM and a 16 GB GPU.
This is the exact modified dealignai UNCENSORED checkpoint converted and quantized to GGUF. It is not derived from the independently modified OrcaRouter model.
File
| File | Size | SHA-256 |
|---|---|---|
Qwen3.8-Flash-Next-UNCENSORED-IQ2_XXS.gguf | 74,658,862,272 bytes (69.531 GiB) | c04bb625ee9a50aaf25d61d2cea8187934117f82bd41aa48918a0f03b37d4173 |
The unquantized BF16 GGUF is available separately at BattleGhost/Qwen3.8-Flash-Next-UNCENSORED-BF16-GGUF.
Conversion and quantization
- Source revision:
2d9a479fd8643c394f1602b4ef1c8757dda4c518 - llama.cpp revision:
c841aeeb8bb2fe417038dadfa9b007cf1a9ef950(build b10687) - GGUF architecture:
qwen4exp - Tensor count: 1,224
- Quantized size: 71,189.73 MiB
- Average: 3.37 BPW
- Main type: IQ2_XXS with an importance matrix from
unsloth/Qwen3.8-Flash-Next-GGUF output_hc_down.weight: IQ4_NLtoken_embd.weight: Q4_K- The large PLE table uses a compatible fallback type because its row geometry cannot use IQ2_XXS directly.
- 195 of 1,224 tensors required compatible fallback types based on tensor geometry.
The source checkpoint stores its PLE n-gram FP8 shards with one shared ngram_embedding.weight_scale. That scale was applied while producing the BF16 intermediate. Quantization did not restore or substitute the original Qwen weights.
Compatibility
Use a recent llama.cpp build containing Qwen3.8 / qwen4exp support. The relevant implementation was merged in ggml-org/llama.cpp#27742.
License and safety
This is a format conversion and quantization of the dealignai checkpoint and remains governed by the Qwen Community License 1.0. Review the source model's license, model card, and disclaimer before use.
The source model has safety refusals removed. Users are responsible for complying with applicable laws and platform policies.