Qwen3.5-4B-3D-Coder-MTP-Repacked

Qwen3.5-4B-3D-Coder-MTP-Repacked

A repacked copy of ccore's Qwen3.5-4B-3D-Coder, a small model tuned to write coding answers and Three.js 3D scenes, with GGUF files that support multi-token prediction.

Made by prithivMLmods 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: Qwen3.5-4B-3D-Coder-MTP-Repacked is a repack by prithivMLmods of Qwen3.5-4B-3D-Coder. ccore made that model as an early preview, tuned from Qwen3.5-4B. It targets general coding and interactive 3D scene generation, written as self-contained Three.js pages in HTML and JavaScript.

The repack carries multi-token prediction, a form of speculative decoding built into the weights. The repository holds safetensors and a set of GGUF files in several quant types.

Who it suits: It suits developers and hobbyists who want a small local model that writes Three.js scenes or general code. The card calls it an early preview, so treat results as drafts.

Industry | Who uses it | How they use it | Example
Software | Web developers | Draft Three.js scenes with a local model | A spinning cube page from a short prompt
Education | Students | Learn 3D on the web by editing generated code | Changing the lights in a scene
Games and media | Creative coders | Sketch interactive 3D ideas quickly | A simple scene with moving shapes
Research | Hobbyist model tinkerers | Try speculative decoding on a 4B model | Comparing speed with and without multi-token prediction

Skip it if you need a model with published benchmark results, since the card gives none.

Where to use it:

• Local apps that load GGUF files, such as LM Studio
• Python with the transformers library, for the safetensors files
• Windows, macOS or Linux, with Python for the download step

Use case | Who | What they get back
3D scene generation | Web developers | A self-contained HTML page with Three.js code
General coding help | Developers | Code and explanations from a 4B model
Faster local runs | Local model users | Speculative decoding through multi-token prediction

How to set it up: You need a few things in place first.

• Python and pip, for the download command
• A local app that can load GGUF files, or transformers for the safetensors files
• Disk space for the file you choose, from 2.32 GB to 8.67 GB for the GGUF set

1. Install the Hugging Face hub package with pip.
2. Download the repository with the hf command.
3. Pick a GGUF quant that fits your memory and load it in your local app. In LM Studio, turn multi-token prediction off, as the card advises.

pip install -U huggingface_hub
hf download prithivMLmods/Qwen3.5-4B-3D-Coder-MTP-Repacked

How to use it: You ask the model for a 3D scene or for code and read what it returns. The card ships recommended sampling settings and a system prompt aimed at Three.js, so check the repository files for them. The model card lists the quant types below.

Quant type | Size | Note
Q3_K_M | 2.32 GB | Low quality, the smallest option listed
Q4_K_S | 2.63 GB | Slightly lower quality, saves space, recommended
Q4_K_M | 2.78 GB | Good quality, the default for most uses, recommended
Q5_K_M | 3.16 GB | High quality, recommended
Q6_K | 3.56 GB | Higher quality still, recommended
Q8_0 | 4.61 GB | The largest quant, usually more than you need
BF16 | 8.67 GB | Full weights, the highest quality

Example: You ask for a Three.js page with a rotating cube and a light that you can move. The model returns a complete HTML file with the JavaScript inside. You save it and open it in a browser. The card warns that some outputs need a retry or a small fix.

For newcomers and experts: Start with Q4_K_M, which the card names as the default for most uses. Ask for one small scene, save the code to a file and open it in a browser to see the result.

Training data came from ccore's Three.js 3D Scenes set, which has 4,146 examples under the MIT license. Each pairs a prompt with its reasoning, the full code and a complexity label. The card also mentions a checkpoint from October 27 that is said to do better on first attempts and on scene quality. Multi-token prediction adds extra output heads that guess several tokens per pass, so no separate draft model is needed.

Title: GGUF file sizes listed in the card
Q3_K_M: 2.32 GB
Q4_K_S: 2.63 GB
Q4_K_M: 2.78 GB
Q5_K_M: 3.16 GB
Q6_K: 3.56 GB
Q8_0: 4.61 GB
BF16: 8.67 GB

Download a GGUF quant with hf
Load it in a local app
Ask for a 3D scene or code
Save the output and open it

Good to know:

• License: Apache 2.0
• Cost to run: none named
• Data it reaches: it downloads files from Hugging Face onto your computer
• Not for: relying on its code unchecked, since the card calls it an early preview and says some outputs need retries or small corrections

On Hugging Face, 13 November 2026

• Task: image text to text
• Library: transformers
• Parameters: 4.7B
• Weights: about 9.3 GB in F16; about 2.3 GB at 4-bit (estimate)
• Formats: Safetensors, GGUF
• Downloads in the last 30 days: 6,042
• Likes: 16
• License: apache-2.0
• Languages: en
• Based on: ccore/Qwen3.5-4B-3D-Coder
• Tags: transformers, safetensors, gguf, qwen3_5, image-text-to-text, llama-cpp, code, threejs, 3d, preview, coder, scenes

Links

• Model: https://huggingface.co/prithivMLmods/Qwen3.5-4B-3D-Coder-MTP-Repacked
• Files: https://huggingface.co/prithivMLmods/Qwen3.5-4B-3D-Coder-MTP-Repacked/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download prithivMLmods/Qwen3.5-4B-3D-Coder-MTP-Repacked

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