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
- qwen3.8
- swift-1.5
- uncensored
- abliterated
- mtp
- llamampere
- ai-plus
qwen3.8-27b-Swift1.5-Uncensored-ATX-AI4
Swift 1.5 Qwen3.8-27B Uncensored in llamAmpere's AI+ format (about 4.2 bits per weight, 14.4 GB). It keeps the MTP draft head. It does not contain the vision tower.
Needs llamAmpere v0.6 or newer. It does not load in mainline llama.cpp. For a file that does, use the IQ4_XS-M quant.
Results
Measured on an RTX 3090 Ti, temperature 1.0, real 100K-token coding, agent and document-question conversations, at least 5,120 tokens written, same llamAmpere v0.6 build for both files:
| ATX-AI4 (this file) | ATX IQ4_XS-M | |
|---|---|---|
| File size | 14.4 GB | 15.6 GB |
| KL divergence vs BF16 (lower = closer to the full model) | 10.4% lower (ratio 0.896, 95% CI 0.860-0.932, 32,400 positions, two reference sets) | baseline |
| 100K prompt read | 1,337 tok/s (+8.9%) | 1,228 tok/s |
| Writing speed | 111.5 tok/s (+2.3%, not significant) | 109.0 tok/s |
| Peak card memory | 17.3 GB (-1.1 GB) | 18.5 GB |
Against llamAmpere v0.5 as published (IQ4_XS-M), v0.6 with this file writes 13.5% faster (112.2 vs 98.8 tok/s, p = 0.019) and reads the 100K prompt 23% faster.
Run it with llamAmpere
One 24 GB card (RTX 3090 / 3090 Ti), 262,144-token context:
git clone -b v0.6 https://github.com/JakeATX/llamAmpere.git
cd llamAmpere
cmake -S . -B build-sm86 -DCMAKE_BUILD_TYPE=Release -DGGML_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES=86 \
-DLLAMA_BUILD_BORINGSSL=ON
cmake --build build-sm86 -j8 --target llama-server
./build-sm86/bin/llama-server \
--hf-repo jakeatx/qwen3.8-27b-Swift1.5-Uncensored-ATX-AI4 \
--hf-file qwen3.8-27b-Swift1.5-Uncensored-ATX-AI4.gguf -c 262144 \
-ngl 99 -fa on -ctk sjkvarn4 -ctv sjkvarn4 -b 4096 -ub 1024 --parallel 1 \
--spec-type draft-mtp-adaptive --spec-draft-n-min-adaptive 3 --spec-draft-n-max 5 --spec-draft-p-min 0
The server listens on http://127.0.0.1:8080 (OpenAI-compatible API). This uses mb-MTP drafting (on by default in v0.6) with an adaptive draft depth of 3 to 5; add --no-spec-mb-mtp for exact sampling.
File
| File | Details |
|---|---|
qwen3.8-27b-Swift1.5-Uncensored-ATX-AI4.gguf | 13.43 GiB (14,416,419,520 bytes), about 4.2 bits per weight, AI+ format 2 (mixed 3-, 4- and 5-bit AI+ types), SHA-256 1106accccb165c6726e46c811955586c462cd8f99879cd0b27c4711ce59744b3 |
Older AI+ files of this model (ATX-Swift-1.5-Qwen3.8-27B-Uncensored-AIplus-4.2bpw.gguf) are retired and do not load in v0.6.
Source and license
The source derives from UkisAI's Swift 1.5 (source revision 15165fce17cb716934a2b15e746d9c7b061d4a0f), a Qwen3.8-27B derivative. This is an uncensored, refusal-reduced derivative (see ABLITERATION.json); the source author reports structural validation but has not published behavioral or capability measurements for it. Licensing: Swift Open License v1.0 for UkisAI's contribution (LICENSE) and Apache 2.0 for Qwen3.8-27B (LICENSE-APACHE-2.0); see NOTICE.