license: other
license_name: swift-open-license-1.0
base_model: d0xin/Swift-1.5-Qwen3.8-27B-Uncensored-BF16
base_model_relation: quantized
library_name: llama.cpp
pipeline_tag: text-generation
tags:
- gguf
- quantized
- imatrix
- qwen3.8
- swift-1.5
- uncensored
- abliterated
- mtp
Swift 1.5 Qwen3.8-27B Uncensored — ATX IQ4_XS-M
This is the Swift 1.5 uncensored BF16 checkpoint quantized with the same per-tensor XS-M map as ATX Qwen3.8-27B IQ4_XS-M and the earlier Swift uncensored build. The source derives from UkisAI's Swift 1.5 and retains its MTP head. Source revision: 15165fce17cb716934a2b15e746d9c7b061d4a0f.
The GGUF contains the text model and MTP draft layer. It does not contain the vision tower. This is an uncensored, refusal-reduced derivative; the source author reports structural validation but has not published behavioral or capability measurements for this derivative. No quality or speed benchmark is claimed for this quant.
| Artifact | Details |
|---|---|
ATX-Swift-1.5-Qwen3.8-27B-Uncensored-IQ4_XS-M.gguf |
14.52 GiB (15,588,550,976 bytes), 4.56 bits per weight, SHA-256 b5c43ac588ef45143afa25441d722ce86d199b77b7d7f808e29a7cffc4f2f244 |
tensor_types_ATX-4-XS.txt |
Exact per-tensor type map, SHA-256 c1b4daffc5d0f623d259179c375f08dab6b516b1a5d92a64a7e71c4812963268 |
imatrix_swift1_unc_transfer.gguf |
Transferred importance matrix from the earlier Swift 1.0 uncensored quant, SHA-256 ef52688ab733e4efa16c6486b0941e11883d15434b4553cfa5e6ae64ed3f3ef1 |
Quantization recipe
The map uses IQ4_XS for bulk weights, Q5_0 for selected attention output, GDN output, and FFN down weights, Q6_K for the output head and selected K/V projections, Q8_0 for selected K/V projections and small GDN vectors, Q5_0 for the MTP layer, and Q4_K for the token embedding. The output tensor types are checked against the base ATX Qwen3.8-27B quant.
This CPU build uses the earlier Swift uncensored model's importance matrix, originally computed from about 226K calibration tokens at Q8_0 precision. It is a transfer matrix, not a calibration measured on Swift 1.5. That choice keeps CPU load and runtime lower while preserving the exact XS-M tensor strategy; any quality effect of the matrix transfer has not been measured.
llama-quantize --imatrix imatrix_swift1_unc_transfer.gguf \
--tensor-type-file tensor_types_ATX-4-XS.txt \
--token-embedding-type q4_K \
Swift-1.5-Qwen3.8-27B-Uncensored-BF16.gguf \
ATX-Swift-1.5-Qwen3.8-27B-Uncensored-IQ4_XS-M.gguf iq4_xs 2
The BF16 conversion and quantization used CPU cores 12–13 at low process priority. Source BF16 weights are not redistributed here. The GGUF's SHA-256 is listed above. A structural tensor check was run before upload. The full-file SHA-256 recorded by Hugging Face Xet was checked against the published file after transfer. Inference quality was not re-evaluated.
License and credits
Distributed under the Swift Open License v1.0, inherited from Swift 1.5. See LICENSE, LICENSE-APACHE-2.0, and NOTICE for the applicable terms and attribution. Credit goes to UkisAI for Swift 1.5, d0xin for the uncensored derivative, and the Qwen team for Qwen3.8-27B.