Multi-Token Prediction (MTP) donor file for Qwen3.5-122B-A10B.
Edited: 2026/5/17 Since llama.cpp has already merged the MTP branch, this model is no longer relevant (although it still works). Please DO NOT USE this model ANYMORE
This donor is ONLY compatible with Qwen3.5-122B-A10B.
Do NOT use with other model sizes (35B, 27B, etc.) or architectures.
What is this?
A lightweight (~1.63 GB) GGUF file containing only the MTP (Multi-Token Prediction) layer tensors extracted from a full Qwen3.5-122B-A10B model. Instead of downloading the full model (tens of GB) just to get MTP support, you can download this donor and inject it into your existing GGUF.
About the script
The convert.py script in this repo works with any GGUF that contains MTP layers. You can use it to extract and share MTP donors for other models too:
bash
1# Extract MTP from any model2python convert.py extract any-model-with-mtp.gguf donor.gguf
34# Merge donor into base model5python convert.py merge base.gguf donor.gguf output.gguf
It has been tested with Qwen3.5-35B, Qwen3.6-27B, and Qwen3.5-122B-A10B — all work the same way. The donor preserves the exact quantization type and per-row metadata of the original tensors.
Usage
Prerequisites
You need a custom build of llama.cpp with MTP support:
