Qwen3.6-14B-A3B-FableVibes

This model started as Qwen3.6-35B-A3B-heretic and was pruned via REAP down to ~14B with ~3B active parameters, removing over half its expert capacity. A single QLoRA pass was then orchestrated entirely by an autonomous AI agent (Steve), utilizing ~4,600 raw reasoning traces from Claude Fable 5 (Mythos-class) to recover capabilities lost during pruning.

Rather than focusing strictly on agentic orchestration, this model serves as a general-purpose reasoning distill. The Fable CoT traces provide structured multi-step reasoning patterns from a frontier-class model, distilled into a footprint that can run on consumer hardware. The Fable traces are further supplemented by Claude Opus reasoning, Qwen tool-calling data, and Evol-Instruct-Code.

Background & Training

• The pruned base was not pre-fine-tuned before this run—the Fable LoRA was applied directly to the REAP output.
• The model leverages Qwen's thinking format and will natively output its step-by-step logic inside <think>...</think> XML blocks (often mapped to the reasoning_content API field).
• Give it a sufficient generation budget. The reasoning pass typically uses hundreds to thousands of tokens to explore state spaces before answering.

Usage

This repository contains the merged bfloat16 base weights. If you are looking for quantized GGUF versions to run in llama.cpp, Ollama, LM Studio, etc., please see the GGUF Repository.