Khazar AI's Qwen3.8 Coding Fine-Tune Hits 870K Downloads as Local GGUF Builds Drop
A 27B coding model distilled on 195M tokens of agentic debugging traces lands in GGUF form, runnable on a single workstation GPU.
- Qwen3.8-27B-Fable-5-Coding-Distilled is a coding fine-tune distilled on 195M tokens of agentic debugging traces.
- Trained to inspect real files before proposing fixes, avoiding speculative guess-and-write patches.
- mradermacher published GGUF quants from 11 GB Q2_K up to 29 GB Q8_0.
- Apache 2.0 license inherited from the Qwen base model, commercial use permitted.
- Runs locally via llama.cpp, Ollama, LM Studio, Jan, and Docker Model Runner.
- Author flags the model card as partially unverified; validate provenance and evals before production.
A new GGUF release of Qwen3.8-27B-Fable-5-Coding-Distilled has appeared on Hugging Face’s trending list. At publication, the repository showed more than 870,000 downloads, though Hugging Face counters represent file requests rather than unique users or installations.
Khazar AI created the coding fine-tune from the dense Qwen3.8-27B base, while mradermacher converted it into llama.cpp-compatible files ranging from an 11 GB Q2_KQ8_0
Why the model looks before it edits
Khazar AI designed the fine-tune for coding agents that inspect project state before proposing changes. Its training data rewards reading files, test harnesses, and CI configuration; checking assumptions against tool output; and preserving existing interfaces or output formats.
Training used 195 million tokens from the fable-5-coding-and-debugging-traces dataset. Distillation in this context means training the model on generated reasoning traces that demonstrate how to investigate software-engineering and debugging tasks. The traces emphasize three habits:
- Inspect the project: Read relevant files, tests, and configuration before editing code.
- Respect constraints: Preserve required output formats, interfaces, and surrounding implementation choices.
- Ground the reasoning: Base conclusions on repository state and tool results instead of speculative explanations.
Repository inspection depends on the agent host exposing file, shell, test, and patch tools, then returning their results to the model. Loading the GGUF in a chat interface alone does not grant access to a codebase.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
