Qwen3.5-397B-A17B-Uncensored-REAP35 GGUF

Expert-pruned version of Qwen3.5-397B-A17B-Uncensored using REAP (Router-weighted Expert Activation Pruning). 35% of experts removed, reducing from 512 to 332 experts per layer.

Based on the uncensored fine-tune by timteh673.

Available Quants

File	Quant	Size	Description
Qwen3.5-397B-A17B-Uncensored-REAP35-Q8_0	Q8_0	259 GB	Full precision pruned, for re-quantization

More quant variants coming soon.

REAP Pruning Method

REAP scores each expert using imatrix calibration data and uniformly removes the lowest-scoring experts from every MoE layer.

Each expert receives a score based on two signals captured during calibration inference:

1. Activation count — how many times the expert was selected by the router
2. Activation magnitude — sum of squared input activations when the expert was active

The final score is: normalized_count x normalized_magnitude

• Base model: Qwen3.5-397B-A17B-Uncensored (512 experts, 10 active per layer)
• Pruned: 512 → 332 experts per layer (35% removed)
• Calibration data: unsloth calibration dataset (base model imatrix)