license: other
license_name: unknown-source-license
tags:
- local-code
- qwen
- uncensored
- quantized
Qwen 3.8 27B Uncensored for Local Code (4-bit S)
An unofficial conversion of Qwen 3.8 27B Uncensored to the .ckpt format used by the Local
Code app. It isn't made or supported by Qwen or by the Local Code developer.
Important
- Reduced safety behavior: this model is based on an uncensored build, so expect fewer
refusals. It can produce harmful or inaccurate content, and you're responsible for how you
use it. - Tool access: in Local Code, models can use tools on your Mac. Review what it does
before approving actions, especially with this model. - It replaces a stock file: the app shows it as Qwen 3.8 27B (4-bit S). An app update
or re-download can overwrite it.
Source and conversion
- Text weights: an uncensored Qwen 3.8 27B, quantized to 4-bit with oMLX, then re-encoded
for Local Code. Quantizing twice loses a little precision. - Original source: not recorded, so the exact upstream model and its license are unknown.
If you recognize it or need clear licensing, check before relying on this model. - Image-reading weights: copied unchanged from Local Code's stock Qwen 3.8 27B model.
- Converter: in the
converterfolder (MIT license)
Install
You need Local Code on an Apple Silicon Mac with Qwen 3.8 27B (4-bit S) already downloaded.
With the script
- Download
qwen_3.8_27b_i4x.ckpt(about 21 GB) andconverter/install-model.sh. - Quit Local Code.
- Run
zsh install-model.sh path/to/qwen_3.8_27b_i4x.ckpt. Your original moves to astock-backupfolder next to the script. - Open Local Code and choose Qwen 3.8 27B (4-bit S).
To go back, quit Local Code and run zsh restore-stock.sh from converter, in the same folder you installed from.
Manually
The file must be named exactly qwen_3.8_27b_i4x.ckpt, or the app won't recognize it. Your
Mac needs about 21 GB free for the download, because the original is kept.
In Finder
- Download
qwen_3.8_27b_i4x.ckptfrom this page. - Quit Local Code.
- In Finder, choose Go > Go to Folder, then paste
~/Library/Containers/com.liuliu.local-code/Data/Documents/Modelsand press Return. - Move the existing
qwen_3.8_27b_i4x.ckptout of that folder to somewhere safe, such as a newLocal Code backupfolder in Documents. Keep it, because it's your only copy of the original. - Move the downloaded file into the
Modelsfolder. - Open Local Code and choose Qwen 3.8 27B (4-bit S).
In Terminal
Quit Local Code first. These commands assume the download is in your Downloads folder. If macOS
asks whether Terminal can access another app's data, allow it.
MODELS="$HOME/Library/Containers/com.liuliu.local-code/Data/Documents/Models"
BACKUP="$HOME/Documents/Local Code backup"
mkdir -p "$BACKUP"
mv -n "$MODELS/qwen_3.8_27b_i4x.ckpt" "$BACKUP/"
mv -n "$HOME/Downloads/qwen_3.8_27b_i4x.ckpt" "$MODELS/"
To go back: quit Local Code, move the converted file out of the Models folder, and move
your original back in.
Check the install
To confirm it's installed, check that the file in Local Code's model folder is 21348925440 bytes:
ls -l "$HOME/Library/Containers/com.liuliu.local-code/Data/Documents/Models/"
Testing
- File checks: all 883 text tensors match the source closely (cosine similarity of at
least 0.992), and 441 image-reading tensors are byte-identical to the stock file. - Loading: Local Code loaded the file on an M5 Max with 48 GB of RAM.
- Known weak spot:
mtp.fcconverted less precisely than the rest (18 dB, against 27 dB or
better elsewhere). - Replies: read normally in the publisher's own use. This isn't a formal evaluation.