Qwen3.8 13.5B

Qwen3.8 13.5B

An experiment that prunes Qwen3.8-27B to half its layers and retrains it, giving a 14.7B model for Japanese and English that the author says is not practical yet.

Made by kusuriya-labs on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Qwen3.8 13.5B is a pruning experiment on Qwen3.8-27B by kusuriya-labs. The author cut the 64 layers down to 32, tried several ways of cutting, measured perplexity and kept the best one. The embeddings were not touched, so the real size is 14.7B parameters, not 13.5B.

Right after pruning the output was completely broken. More training followed, with two Qwen3.8 models acting as teachers, and a large part of the quality came back. The model speaks Japanese and English.

Who it suits: It suits researchers and hobbyists who like model pruning and want a starting point for more training. The author states it is not practical, and its answers drift from the question.

Industry | Who uses it | How they use it | Example
Research | ML researchers | Study layer pruning and recovery by distillation | Comparing it with the 27B model
Home labs | Enthusiasts | Try a pruned model and push it further | Running extra training with spare compute
Education | Students of model compression | See what pruning half the layers does | Reading outputs before and after training
Software | Developers | Test a smaller Qwen3.8 variant | Checking Japanese and English replies

Skip it if you need a reliable assistant, since the author says it is not one yet.

Where to use it:

• Python with transformers, for the safetensors files
• Local apps that load the GGUF files
• Windows, macOS or Linux, with Python for the download step

Use case | Who | What they get back
Pruning research | Researchers | A concrete half-depth model to study
Further training | People with spare compute | A base that already speaks Japanese and English
Japanese and English tests | Hobbyists | Replies that may drift from the input

How to set it up: You need a few things in place first.

• Python and pip, for the download command
• A local app for the GGUF files, or transformers for the safetensors files
• About 7.4 GB of disk at 4-bit, by the record's estimate

1. Install the Hugging Face hub package with pip.
2. Download the repository with the hf command.
3. Load the files in your tool. The model card has no run commands.

pip install -U huggingface_hub
hf download kusuriya-labs/Qwen3.8-13.5B

How to use it: You load the model and give it prompts in Japanese or English. Expect replies that are loosely related to the input. The card is clear that this is an experiment, so treat the output as a curiosity.

Example: You load the GGUF file in a local app and ask a short question in Japanese. The model answers in Japanese or English, but the reply may not match the question well. The author says so openly.

For newcomers and experts: Newcomers should start with a bigger, finished model, and come back to this one for curiosity. Try one short prompt to see what a half-depth model does.

The pruning removed half of the 64 layers of Qwen3.8-27B. The author compared several removal patterns by perplexity and chose the best. Training afterward used two teacher models to recover the output. The author invites people with more compute to continue the training and aim for a stronger model.

Title: Layer count, from the card
Qwen3.8-27B: 64 layers
This model: 32 layers

Prune half the layers of Qwen3.8-27B
Measure perplexity and pick the best cut
Train with two teacher models
Publish the recovered model

Good to know:

• License: Apache 2.0
• Cost to run: none named
• Data it reaches: it downloads files from Hugging Face onto your computer
• Not for: practical use, since the author says it is an experiment and its answers drift from the input

On Hugging Face, 13 November 2026

• Task: text generation
• Parameters: 14.7B
• Weights: about 7.4 GB at 4-bit (estimate)
• Formats: Safetensors, GGUF
• Downloads in the last 30 days: 414
• Likes: 12
• License: apache-2.0
• Languages: ja, en
• Based on: Qwen/Qwen3.8-27B
• Tags: safetensors, gguf, qwen3.8, text-generation, conversational

Links

• Model: https://huggingface.co/kusuriya-labs/Qwen3.8-13.5B
• Files: https://huggingface.co/kusuriya-labs/Qwen3.8-13.5B/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download kusuriya-labs/Qwen3.8-13.5B

Then follow the usage in its model card. Nothing in this kit runs by itself.