license: other
license_name: swift-open-license-1.0
license_link: LICENSE
base_model: ajgazin/Swift-1.5-Qwen3.8-27B-Uncensored-MTP
base_model_relation: quantized
library_name: mlx
pipeline_tag: text-generation
language:
- en
- multilingual
tags: - mlx
- qwen3_5
- qwen3_8
- swift
- abliterated
- uncensored
Swift-1.5-Qwen3.8-27B-Uncensored, MLX 4-bit
An MLX 4-bit build (4.5 bits per weight, about 14 GB) of
ajgazin/Swift-1.5-Qwen3.8-27B-Uncensored-MTP,
the uncensored variant of ukisai/Swift-1.5-Qwen3.8-27b
(a Qwen3.8-27B fine-tune by UkisAI).
Modified files: the weights were quantized to MLX format with mlx-lm 0.32.2, standard uniform
4-bit quantization. No other change to the weights or the chat template:
mlx_lm.convert --hf-path ajgazin/Swift-1.5-Qwen3.8-27B-Uncensored-MTP --mlx-path out -q --q-bits 4
Text generation through mlx-lm only; the vision input and the MTP draft head of the source
checkpoint were not tested in this build.
Use
pip install "mlx-lm>=0.32.2"
mlx_lm.generate --model tmthang86/Swift-1.5-Qwen3.8-27B-Uncensored-mlx-4bit --max-tokens 3000 --prompt "..."
It is a thinking model: leave room for the reasoning before the answer.
Checked
Run on an Apple Silicon Mac with 48 GB as a continuity checker on one Vietnamese novel chapter
with 10 seeded errors (3 seeded runs and 1 clean run): 57 % ± 15 of the errors caught, no false
alarms, 3 of 15 calls returned an empty answer after long thinking, about 11 minutes per call.
Licence
Swift Open License 1.0 (UkisAI) for the Swift contribution, Apache 2.0 for the Qwen3.8-27B base
model (Alibaba Cloud), as in the source repository. LICENSE, LICENSE-APACHE-2.0 and NOTICE
are copied unchanged. Commercial use above the licence's revenue threshold needs a licence from
UkisAI: read LICENSE before any commercial use. This build is not made or endorsed by UkisAI.