khazarai

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

README

License: apache-2.0

Model Details

Table

Base model	Qwen/Qwen3.8-27B (dense, ~27B params, Apache 2.0)
Fine-tuning method	Distillation on curated reasoning traces *(SFT / LoRA
Training data size	195M tokens
Domain focus	Coding, debugging, agentic tool-use (shell, test harnesses, CI/lint workflows)
Context length	Inherits base model context window (confirm exact value)
License	Apache 2.0 (inherited from base — confirm)
Languages	English (code + natural language)

Training Data

The model was fine-tuned on 195M tokens of reasoning traces distilled for coding and agentic-debugging scenarios. Traces emphasize:

Reading and verifying project state (files, test harnesses, CI configs) before proposing a fix
Working within existing code/output constraints instead of unnecessary rewrites
Concise, evidence-grounded reasoning over long speculative chains

Intended Use

This model is intended for:

Agentic coding assistants that need to inspect a repository before acting
Debugging and root-cause analysis tasks (e.g., CI failures, lint errors, shell script bugs)
Code review and refactoring within existing project constraints

It is not intended for use as a general-purpose chat assistant without further evaluation.

Example: Grounded vs. Speculative Reasoning

The distillation objective specifically targets the difference below, observed when comparing this fine-tune against the undistilled base model on an agentic bash-debugging task (a CI pipeline failing on a shell script with quoting bugs):

This model (fine-tuned): Immediately inspects the actual project files (test harness, target script) before forming a fix, keeping its response short and grounded in verified state rather than assumptions.
Base model: Reasons at length about what the script might contain, repeatedly acknowledges it hasn't seen the files, and ultimately proposes a fix based on guesses rather than a tool call — risking a rewrite that breaks pinned output formats the task explicitly asked to preserve.

This kind of "check before you fix" behavior is the primary signal the 195M-token distillation set was built to reinforce.

Limitations

Inherits the general limitations of the Qwen3.8-27B base model (knowledge cutoff, potential hallucination on unfamiliar codebases, no execution sandboxing on its own).
Distillation was focused on coding/agentic-debugging traces — general chat and non-coding reasoning quality has not been separately evaluated.
As with any fine-tune, behavior on out-of-distribution tasks (outside shell/CI/test-harness-style debugging) should be validated before production use.

Citation

If you use this model, please cite:

bibtex

@misc{qwen38_27b_coding_distilled,  title  = {Qwen3.8-27B-Coding-Distilled},  author = {khazarai},  year   = {2026},  note   = {Fine-tune of Qwen/Qwen3.8-27B, distilled on 195M tokens of agentic coding reasoning traces},  url    = {https://huggingface.co/khazarai/Qwen3.8-27B-Coding-Distilled}}

Acknowledgements

Built on top of Qwen/Qwen3.8-27B by the Qwen team.

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider

khazarai

Model Tree

Base

Qwen/Qwen3.8-27B

Fine-tuned

this model

Input Modalities

TextImageVideo

Output Modalities

Text

Supported Functionality

Dedicated Endpoints

Explore FriendliAI today

Get started Talk to an engineer

README

License: apache-2.0

Model Details

Table

Base model	Qwen/Qwen3.8-27B (dense, ~27B params, Apache 2.0)
Fine-tuning method	Distillation on curated reasoning traces *(SFT / LoRA
Training data size	195M tokens
Domain focus	Coding, debugging, agentic tool-use (shell, test harnesses, CI/lint workflows)
Context length	Inherits base model context window (confirm exact value)
License	Apache 2.0 (inherited from base — confirm)
Languages	English (code + natural language)

Training Data

The model was fine-tuned on 195M tokens of reasoning traces distilled for coding and agentic-debugging scenarios. Traces emphasize:

Reading and verifying project state (files, test harnesses, CI configs) before proposing a fix
Working within existing code/output constraints instead of unnecessary rewrites
Concise, evidence-grounded reasoning over long speculative chains

Intended Use

This model is intended for:

Agentic coding assistants that need to inspect a repository before acting
Debugging and root-cause analysis tasks (e.g., CI failures, lint errors, shell script bugs)
Code review and refactoring within existing project constraints

It is not intended for use as a general-purpose chat assistant without further evaluation.

Example: Grounded vs. Speculative Reasoning

This model (fine-tuned): Immediately inspects the actual project files (test harness, target script) before forming a fix, keeping its response short and grounded in verified state rather than assumptions.
Base model: Reasons at length about what the script might contain, repeatedly acknowledges it hasn't seen the files, and ultimately proposes a fix based on guesses rather than a tool call — risking a rewrite that breaks pinned output formats the task explicitly asked to preserve.

This kind of "check before you fix" behavior is the primary signal the 195M-token distillation set was built to reinforce.

Limitations

Inherits the general limitations of the Qwen3.8-27B base model (knowledge cutoff, potential hallucination on unfamiliar codebases, no execution sandboxing on its own).
Distillation was focused on coding/agentic-debugging traces — general chat and non-coding reasoning quality has not been separately evaluated.
As with any fine-tune, behavior on out-of-distribution tasks (outside shell/CI/test-harness-style debugging) should be validated before production use.

Citation

If you use this model, please cite:

bibtex

@misc{qwen38_27b_coding_distilled,  title  = {Qwen3.8-27B-Coding-Distilled},  author = {khazarai},  year   = {2026},  note   = {Fine-tune of Qwen/Qwen3.8-27B, distilled on 195M tokens of agentic coding reasoning traces},  url    = {https://huggingface.co/khazarai/Qwen3.8-27B-Coding-Distilled}}

@misc{qwen38_27b_coding_distilled,  title  = {Qwen3.8-27B-Coding-Distilled},  author = {khazarai},  year   = {2026},  note   = {Fine-tune of Qwen/Qwen3.8-27B, distilled on 195M tokens of agentic coding reasoning traces},  url    = {https://huggingface.co/khazarai/Qwen3.8-27B-Coding-Distilled}}

Acknowledgements

Built on top of Qwen/Qwen3.8-27B by the Qwen team.