Qwen3.8-27B-Fable-5-Coding-Distilled

Qwen3.8-27B-Fable-5-Coding-Distilled

Debugs and reviews code by checking real files before it proposes a fix, a Qwen3.8-27B fine-tune trained on 195M tokens of reasoning traces. For developers who build coding agents.

Made by khazarai 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: This repository holds Qwen3.8-27B after further training on 195M tokens of distilled reasoning traces about software engineering and debugging, with a focus on coding and agent work. It aims for grounded, tool-using reasoning. It should read the real files and check its assumptions, instead of guessing and then writing code.

Who it suits: It suits developers who build coding assistants that look inside a repository before they act, and teams who debug CI and lint problems.

Industry | Who uses it | How they use it | Example
Software | Developers | Power a coding agent that inspects a project first | Find why a build script breaks
Operations | DevOps teams | Debug CI failures, lint errors and shell script bugs | Fix a failing pipeline step
Quality | Reviewers | Review and refactor code within existing project limits | Suggest a small fix, not a rewrite
Research | ML researchers | Study distillation for grounded agent behavior | Compare it with the base model on a debugging task

If you want a general chat assistant, you can skip it, because the card says it was not tested for that.

Where to use it:

• Agentic coding setups that give the model access to files and tools
• The transformers library, which Hugging Face lists for it
• Windows, macOS or Linux, where you download the files with Python

Use case | Who | What they get back
Debugging | Developers | A short, evidence-based answer after reading the files
Code review | Reviewers | Suggestions that keep existing output formats
Root-cause analysis | DevOps teams | A diagnosis based on the real project state

How to set it up: You need Python with pip and a machine with enough memory for a 27.8B model. The card gives no run code, so follow the base model's own page for loading.

• Python and pip
• The transformers library
• About 55.6 GB of disk for the BF16 weights, or about 13.9 GB at 4-bit, which is an estimate

1. Install the Hugging Face client with pip.
2. Download the repository with the hf command.
3. Load it with transformers, as the base model's page describes.

pip install -U huggingface_hub
hf download khazarai/Qwen3.8-27B-Fable-5-Coding-Distilled

How to use it: Give it a coding or debugging task in an environment where it can read files and run checks. The card shows no sample code, so follow the base model's instructions for prompts and settings.

Behavior | This fine-tune | Base model
First move | Inspects the actual project files | Reasons at length about what a script might hold
Answer style | Short, grounded in verified state | Builds a fix from guesses
Risk | Keeps output formats the task asked to preserve | May rewrite and break those formats

Example: A CI pipeline fails on a shell script with quoting bugs. The card says this model first inspects the test setup and the script itself, then proposes a short fix based on what it saw.

For newcomers and experts: Think of it as a coding helper that opens the files before it answers. The first small thing to try is a debugging task where it can read your project.

The card says the 195M-token set was built to reward this check-before-fixing habit. It stresses reading project state, working inside existing constraints, and short reasoning backed by evidence. The fine-tuning method is cut off in the card's table, so it is not clear whether it was SFT or LoRA.

Qwen3.8-27B base model
Distilled on 195M tokens of coding reasoning traces
Coding and debugging fine-tune
Agent reads files, then proposes a fix

Good to know: Size: Hugging Face lists 27.8B parameters, and the card says about 27B.

Title: Weights in GB
BF16: 55.6 GB
4-bit estimate: 13.9 GB

• Unconfirmed details: the card itself says to verify the license inheritance, the dataset sources, the training settings, the hardware and the exact context length.
• Images: Hugging Face tags it as image text to text, but the card describes only coding work and says nothing about pictures.
• Other versions: 6 quantized and 1 fine-tuned version made by other people.
• Limits: it inherits the base model's limits, and its quality on general chat and non-coding reasoning was not tested.
• License: Apache 2.0, as the License line says. The card says it follows the base model, and asks you to confirm.
• Cost to run: no API key, account or paid service is named.
• Data it reaches: the prompts and files you give it, and Hugging Face for the download.
• Not for: general chat use without further testing, which the card warns against, and untested tasks in production.

On Hugging Face, 13 November 2026

• Task: image text to text
• Library: transformers
• Parameters: 27.8B
• Weights: about 55.6 GB in BF16; about 13.9 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 471
• Likes: 11
• License: apache-2.0
• Languages: en
• Based on: Qwen/Qwen3.8-27B
• Quantized versions: 6
• Fine-tunes: 1
• Tags: transformers, safetensors, qwen3_5, image-text-to-text, coding, agentic-coding, reasoning, distilled, unsloth, conversational

Links

• Model: https://huggingface.co/khazarai/Qwen3.8-27B-Fable-5-Coding-Distilled
• Files: https://huggingface.co/khazarai/Qwen3.8-27B-Fable-5-Coding-Distilled/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download khazarai/Qwen3.8-27B-Fable-5-Coding-Distilled

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