Qwen3.6-27B-MTP-GGUF

Qwen3.6-27B-MTP-GGUF

Codes and reasons with Qwen3.6 27B in GGUF files by Unsloth that carry multi-token prediction, which the card says speeds up replies about 1.5 to 2 times with no accuracy loss. Made for coding agents.

Made by unsloth on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: This repository holds GGUF versions of Qwen3.6-27B, a 27.3B parameter model that reads text and images and focuses on coding. Unsloth made the files with multi-token prediction, or MTP, included. The 4-bit weights are estimated at about 13.7 GB.

Who it suits: Developers who want a strong local coding model and faster replies on a good GPU. The card says it handles frontend work and reasoning across a whole repository.

Industry | Who uses it | How they use it | Example
Software | Developers | Use a local model for agentic coding | Fix a bug across several files
Software | Frontend teams | Build and adjust interface code | Draft a page layout from a description
Research | AI researchers | Study multi-token prediction in llama.cpp | Time replies with MTP on and off
Self-hosting | Privacy minded teams | Keep code on their own machine | Run the model inside a private network

People with a small laptop can skip it, because a 27B model needs a lot of memory.

Where to use it:

• llama-server from a llama.cpp build, as in the card
• Unsloth Studio, which the card says sets the MTP options for your hardware
• Coding tools such as Codex and OpenCode, which the card says now work with the developer role
• Windows, macOS or Linux with Python, for downloading the files

Use case | Who | What they get back
Agentic coding | Developers | Code changes and repository level reasoning
Faster replies | Anyone with a GPU | The same answers with fewer waiting steps
Tool calling | Agent builders | Better handling of nested objects in tool calls
Image input | Developers | Answers about pictures you send

How to set it up:

• Python, for the Hugging Face hub tools
• A llama.cpp build that includes llama-server, built with CUDA on for a GPU
• Disk space and memory for the file you choose, with the 4-bit weights estimated at about 13.7 GB

1. Install the Hugging Face hub tools with pip.
2. Download the files with the hf command, or let llama-server fetch them.
3. Build llama.cpp from its source, as the card's first block shows. Turn CUDA off for a CPU or Metal build.
4. Start llama-server with the card's second block.

pip install -U huggingface_hub
hf download unsloth/Qwen3.6-27B-MTP-GGUF

How to use it: Start the server with the command below. It fetches the UD-Q4_K_XL file, offloads layers to the GPU, sets a context of 8192, turns on flash attention and enables the MTP drafter. The card says more than one parallel slot and the mmproj option are not yet supported with MTP.

export LLAMA_CACHE="unsloth/Qwen3.6-27B-MTP-GGUF"
./llama.cpp/llama-server \
    -hf unsloth/Qwen3.6-27B-MTP-GGUF:UD-Q4_K_XL \
    -ngl 99 -c 8192 -fa on -np 1 \
    --spec-type draft-mtp --spec-draft-n-max 2

Tool | What it does
-hf | Loads the model from Hugging Face, with the file type after the colon
-ngl 99 | Offloads layers to the GPU
-c 8192 | Sets the context size
-np 1 | Uses one parallel slot, the only value MTP supports for now
--spec-type draft-mtp | Turns on the multi-token prediction drafter
--spec-draft-n-max 2 | Allows up to 2 drafted tokens per step

Example: You start the server with the card's command and ask the model to fix a bug in a small project. The drafter guesses a couple of tokens ahead, and the main model checks them. The card says this makes replies about 1.5 to 2 times faster with no accuracy loss.

For newcomers and experts: Run the card's command as it is first. Use Unsloth Studio if you would rather not tune options by hand, since it sets the MTP options for your hardware. Add one change at a time.

The card describes two upgrades in Qwen3.6. The first is stronger agentic coding, covering frontend work and reasoning across a repository. The second is thinking preservation, an option that keeps reasoning from earlier messages so iterative work needs less repeated effort.

Start llama-server with the model and MTP
The drafter guesses up to 2 tokens ahead
The main model checks each guess
Confirmed tokens join the reply

Good to know:

• Speed: The card says MTP gives about 1.5 to 2 times faster inference with no accuracy loss.
• Limits: More than one parallel slot and the mmproj option are not yet supported with MTP.
• Context: The window is 262,144 tokens natively and can reach 1,010,000 when extended.
• License: Apache 2.0.
• Cost to run: none named. No API key or paid service appears in the source.
• Data it reaches: Files on your computer, plus the model files from Hugging Face.
• Not for: Running several parallel slots with MTP, for now.

On Hugging Face, 13 November 2026

• Task: image text to text
• Library: transformers
• Parameters: 27.3B
• Weights: about 13.7 GB at 4-bit (estimate)
• Formats: GGUF
• Downloads in the last 30 days: 46,035
• Likes: 1,344
• License: apache-2.0
• Based on: Qwen/Qwen3.6-27B
• Quantized versions: 8
• Fine-tunes: 2
• Tags: transformers, gguf, unsloth, qwen, qwen3_5, image-text-to-text, imatrix, conversational

Links

• Model: https://huggingface.co/unsloth/Qwen3.6-27B-MTP-GGUF
• Files: https://huggingface.co/unsloth/Qwen3.6-27B-MTP-GGUF/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download unsloth/Qwen3.6-27B-MTP-GGUF

Then follow the usage in its model card. Nothing in this kit runs by itself.