MTP-only GGUF subset of Qwen3-Next models
This is a supplement for Qwen3-Next-80B-A3B-Thinking-based models (including fine-tunes and abliterated models) without MTP tensors.
This repository contains an MTP-only subset of Qwen/Qwen3-Next-80B-A3B-Thinking which provides the draft model for speculative decoding in the GGUF format.
It accelerates token generation using speculative decoding with the draft model from the original Qwen model. In most cases, this is sufficient to accelerate Qwen-based derivative models even if this draft model is not trained from them.
Note that however, the performance metrics heavily depend on the derivative model you use, your machine and your MTP settings.
Benchmark it before blindly trusting it.
Using this Model
It can be used in two ways:
- As a separate draft model file (Method 1)
- As a donor for grafting the draft model into a Qwen-based model (Method 2; Recommended)
Method 1: Separate Draft Model File
It is easy to begin with but memory-inefficient as it does not share some tensors with the original model.
If you find the draft model can accelerate a Qwen-based model you use, grafting the draft model (Method 2) is recommended (note: switching to Method 2 may slightly change the acceptance rate).
If you use llama-server, you may configure like this:
llama-server \
--model Qwen3-Next-80B-A3B-Instruct-Q4_K_M.gguf \
--model-draft Qwen3-Next-80B-A3B-Instruct-MTP-ONLY-Q4_K_M.gguf \
... \
--spec-type draft-mtp
--spec-draft-n-max 4
--model specifies the original Qwen-based model and
new --model-draft specifies a file from this repository.
You also need --spec-type draft-mtp to enable the draft model.
Once the draft model is enabled, you may configure the rest of MTP options
as you like (in this example, custom --spec-draft-n-max is specified).
Method 2: Grafting the Draft Model (Using this as a Donor)
This is recommended.
First, download convert.py
written by @buzz to transplant MTP tensors.
You may also need to install some dependencies required by this script.
Then, you can run this script like:
# ./convert.py INPUT MTP OUTPUT
python3 convert.py \
Qwen3-Next-80B-A3B-Instruct-Q4_K_M.gguf \
Qwen3-Next-80B-A3B-Instruct-MTP-ONLY-Q4_K_M.gguf \
Qwen3-Next-80B-A3B-Instruct-Q4_K_M+MTP.gguf
The second argument of convert.py is a GGUF file (donor) from this repository.
Once the grafted GGUF file is created, you may use this like a regular Qwen model with embedded draft model.
This is an example for llama-server users.
llama-server \
--model Qwen3-Next-80B-A3B-Instruct-Q4_K_M+MTP.gguf \
... \
--spec-type draft-mtp
--spec-draft-n-max 4
The output of convert.py must be specified as the model file name
(DO NOT use --model-draft in this case).
--spec-type draft-mtp enables the draft model transplanted into the main one.
Additional Quantization (Q4_K_M, Q5_K_M, Q6_K and Q8_0)
Quantized GGUF files are provided so that deploying the draft model easier.
As no Unsloth MTP quants for Qwen3-Next is released,
tensor types are automatically chosen by llama-quantize.
It is not required to match the quantization level.
For instance, you may pair Q6_K-quantized draft model with
the IQ4_NL-quantized main model.
Conversion Process
The --mtp option of convert_hf_to_gguf.py is the crucial part of this
conversion process because this option does exactly what the author expects:
create an MTP-only GGUF subset.
For additional quantization, the llama-quantize tool (llama.cpp)
is used without options.
License and Copyright
For all GGUF files under this repository, the license terms of the original Qwen model (Apache License version 2.0) applies (as the author of this repository did not perform any changes significant enough for own copyright):
Copyright 2025 Alibaba Cloud
No NOTICE files are attached in the original model.
This README file is licensed under the terms of CC-BY-4.0.
Copyright 2026 a4lg.
Links: Sources
- The Original Model: Qwen/Qwen3-Next-80B-A3B-Thinking