Qwen3.6-35B-REAP-MTP-UD-GGUF-Collection

Qwen 3.6 REAP-Pruned MTP GGUF Collection

This archive contains a collection of pruned models derived from Unsloth's Qwen 3.6 35B MTP GGUF. These models have been pruned using the REAP (Routed-Expert Pruning) strategy while carefully maintaining the Multi-Token Prediction (MTP) layer (Layer 40) functionality.

Models Included

1. ATBender Configuration (192 Experts)

Pruned based on the atbender/Qwen3.6-VL-REAP-26B-A3B strategy.

• Qwen3.6-35B-A3B-UD-IQ3_S-REAP.gguf
• Qwen3.6-35B-A3B-UD-IQ3_XXS-REAP.gguf
• Qwen3.6-35B-A3B-UD-Q3_K_M-REAP.gguf
• Qwen3.6-35B-A3B-UD-Q3_K_XL-REAP.gguf

2. RangerX Configuration (180 Experts, Ratio 0.3)

Pruned based on the RangerX/Qwen3.6-35B-REAP-Pruned-ratio-0.3 strategy (reverse-engineered from router weights).

• Qwen3.6-35B-A3B-UD-IQ3_S-REAP-RangerX.gguf
• Qwen3.6-35B-A3B-UD-IQ3_XXS-REAP-RangerX.gguf
• Qwen3.6-35B-A3B-UD-Q3_K_M-REAP-RangerX.gguf
• Qwen3.6-35B-A3B-UD-Q3_K_XL-REAP-RangerX.gguf

Creation Process

These models were created using a custom Python script that:

1. Identified Kept Experts: For the ATBender set, indices were pulled from reap_metadata.json. For the RangerX set, indices were reverse-engineered by comparing the pruned router gate weights against the original unpruned gates via cosine similarity.
2. Surgical Slicing: The script directly sliced the GGUF tensors along the expert dimension (dim0 in the raw memory layout) without dequantizing the individual blocks.
3. MTP Preservation: All MTP-specific tensors (Layer 40) were handled correctly. For tensors with expert dimensions in Layer 40, the Layer 39 pruning strategy was applied to maintain architectural consistency.
4. Metadata Updates: The qwen35moe.expert_count was updated in the GGUF headers to ensure compatibility with llama.cpp.

Usage

To run with MTP support, use a recent build of llama.cpp (May 2026 or later) with the following flag: --spec-type draft-mtp --spec-draft-n-max 3

---

Created by Gemini CLI on May 21, 2026.