---

license: other

license_name: qwen-community-license-1.0

base_model:

- ji-farthing/Qwen3.8-Flash-Next-Uncensored-ik-llama-GGUF

base_model_relation: quantized

pipeline_tag: text-generation

tags:

- gguf

- ik_llama

- qwen4exp

- text-only

- abliterated

- uncensored

- text-generation

---

Note: this is a modified GGUF, based on https://huggingface.co/ji-farthing/Qwen3.8-Flash-Next-Uncensored-ik-llama-GGUF

Want to run this with Strata? See Running with Strata below.

Acknowledgments

Huge thanks to ji-farthing — every tensor in this release comes from their excellent ik_llama.cpp IQ3_KT / IQ4_KT quantizations. All credit for the quantization work belongs to them; this is only a repackaging. If you find v4-lite useful, please go check out and follow ji-farthing's models for the original releases.

This release is shared under the same terms as the source: Qwen Community License 1.0 (Copyright (c) 2026 Qwen) — see the License section below.

Why v4-lite: ~2.7 GiB less VRAM than v4 IQ3_KT (approx. 49 GiB vs 52 GiB in the uploader's local test) for about +0.03 mean KLD — details in Why v4-lite.

v4-lite: main model + standalone IQ4_KT PLE GGUF

This directory is a complete two-file split model. Copy both GGUFs to another machine and you carry all model weights; at runtime there is no need for v2, v4, or any earlier expert patches.

| File | Contents | Size |

| --- | --- | --- |

| Qwen3.8-Flash-Next-V4-Lite-IQ3_KT-IQ4PLE-00001-of-00002.gguf | 1,223 tensors: v4 dense and non-expert weights + all v2 IQ3_KT routed experts | 49.52 GiB |

| Qwen3.8-Flash-Next-V4-Lite-IQ3_KT-IQ4PLE-00002-of-00002.gguf | PLE table only, v2 IQ4_KT | 25.03 GiB |

Both files total about 74.55 GiB. The second file is just the standalone PLE GGUF, using standard split naming so that Strata can auto-discover it. Keep both files in the same directory with their matching filenames. SHA256SUMS.txt can be used to verify the files after transfer; manifest.json records each tensor's original source and SHA256 — source filenames are for traceability and are not runtime dependencies.

The retained v4 parts include all 387 high-precision small projections, the shared experts, the output head, the token embedding, the PLE key/value projections and the other non-routed expert tensors. All 144 routed experts across the 48 layers come from v2 IQ3_KT. The PLE table comes from v2 IQ4_KT, type 155, shape [160, 320001536]. All tensors are direct copies of quantized bytes — no re-quantization or calibration was performed.

The first file keeps v4's tokenizer, chat template, architecture parameters and other original metadata encodings. Only the model name and split info were adjusted and a combined description added; the model has 1,224 tensors in total, and the second split holds only the single PLE tensor.

Why v4-lite

Compared with ji-farthing's v4 IQ3_KT, v4-lite needs about 2.7 GiB less VRAM in the uploader's local test (approx. 49 GiB vs approx. 52 GiB), at the cost of roughly +0.03 mean KLD (also measured locally by the uploader). If your GPU is tight on VRAM, this trade may be worth it.

These files require ik_llama.cpp (the IQ3_KT / IQ4_KT tensor types), same as the source v4 release.

Vision (mmproj)

For multimodal use, grab the official mmproj built from this same uncensored source (the vision tower is untouched by abliteration, so it matches these weights):

mmproj-Qwen3.8-Flash-Next-Uncensored-F16.gguf

Download it alongside the two files above and pass it with --mmproj in ik_llama.cpp / llama.cpp. It is hosted upstream by orcarouter, so it is linked here rather than re-uploaded.

Running with Strata

Strata's official releases do not support IQ3_KT (its supported sizes are Q2_0 / IQ2_XS / IQ3_XXS / IQ3_S plus the Coder / Swift / Unsloth variants). To run this model with Strata, use the uploader's fork with ik_llama.cpp support and Vulkan vision:

Amarillys/Strata — feat/ik-support-with-vulkan-vision

Disclaimer

The source model has its safety alignment removed, and repackaging does not restore it. Read the source card's disclaimer before use.

License

Qwen Community License 1.0 — Copyright (c) 2026 Qwen. This is a derivative/repackaging of ji-farthing's quantization release; the original license and copyright notice are retained. See the LICENSE file for the full text.