exnivo/Qwen3.8-20B-Minitron

Qwen3.8-20B-Minitron: A Pruned Qwen Derivative

exnivo/Qwen3.8-20B-Minitron is an experimental, structurally-pruned version of the Qwen/Qwen3.8-27B model. This is not a quantization but a physical reduction of the model's architecture, decreasing its size from 27.357 billion to 19.746 billion parameters by removing 20 decoder layers (from 64 to 44).

Key Characteristics

• Structural Pruning: Achieved a 27.82% parameter reduction through iterative marginal-KL group pruning, removing approximately 7.61 billion parameters.
• Knowledge Distillation Recovery: The pruned model underwent multi-stage recovery using original Qwen3.8-27B teacher logits, top-k knowledge distillation, next-token training, and targeted repair with mixed educational/instruction/math/code data.
• Fully Merged Weights: The uploaded checkpoint contains fully merged weights, requiring no adapters or pruning scripts for loading.
• Internal Evaluation: Custom internal checks showed 81.2% overall accuracy, with 100% in arithmetic and logic, and 75% in Python, though these are not standardized benchmarks.

Limitations

As an experimental compressed model, it exhibits some weaknesses, particularly in pattern/sequence reasoning. It may also produce incorrect reasoning with unusual wording, and some capabilities not covered by the recovery tests might be reduced. The vision tower is retained but was not the focus of the recovery process, and its quality should not be assumed to match the original Qwen3.8-27B.

When to Use This Model

This model is ideal for developers seeking a more efficient, smaller footprint alternative to Qwen3.8-27B for general text generation and understanding tasks, especially when resource constraints are a concern. It's suitable for applications where the trade-off between model size and some advanced reasoning capabilities is acceptable.