---
title: How to Run Qwen3.5-4B GLB 3D Coder Locally and Use It in Godot
url: "https://goremy.ai/blog/articles/qwen3-5-4b-glb-3d-coder-run-locally-godot"
author: Luis Chavez-Mattos
editor: Luis Chavez-Mattos
editor_title: Director, Product
published: "2026-11-04T15:40:06.067Z"
updated: "2026-11-04T15:40:06.067Z"
topic: Open Model Run Guides
description: "Run CCORE's Qwen3.5-4B GLB 3D Coder locally from a GGUF, set the sampling and system prompt, then export Three.js scenes as GLB files for Godot."
---

# How to Run Qwen3.5-4B GLB 3D Coder Locally and Use It in Godot

> Run CCORE's Qwen3.5-4B GLB 3D Coder locally from a GGUF, set the sampling and system prompt, then export Three.js scenes as GLB files for Godot.

Edited by Luis Chavez-Mattos, Director, Product · Published 2026-11-04

## How do you run Qwen3.5-4B GLB 3D Coder locally and get assets into Godot?

Download either the Q4_K_M or Q8_0 GGUF from the ccore/Qwen3.5-4B-GLB-3D-Coder Hugging Face page and load it in a GGUF-capable app such as LM Studio. Paste in the recommended system prompt, set temperature 0.8, top K 40, top P 0.95 and min P 0.0, then ask for an asset in plain English. The model writes a standalone Three.js HTML page with an Export GLB button. Open it in a browser, export the `.glb`, and import it into Godot.

## TL;DR

- The model is an independent **4B coding model from CCORE**, built on Qwen3.5-4B, and it is not an official Qwen release.
- It does not output meshes directly. It writes **one standalone HTML file using Three.js** that renders the asset and exports a GLB.
- Two GGUF files exist, **Q4_K_M and Q8_0**, and the vision projector (`mmproj`) is optional and only needed for image-reference input.
- Use the author's **system prompt and sampling values** (temperature 0.8, top K 40, top P 0.95, min P 0.0) as a starting point.
- If you downloaded a GGUF before **November 2, 2026 at 04:50 UTC**, delete it and download again, because the filenames did not change.
- This is an **early preview**. Some generations need a retry or a small code fix, and English prompts work better.

## What is the Qwen3.5-4B GLB 3D Coder?

It is a small coding model trained to generate Three.js scenes and assets with GLB export, so the results can go into editors such as Godot. CCORE, an independent developer, based it on Qwen/Qwen3.5-4B. The training data spans 25 game-asset domains, including characters, enemies, weapons, props, modular architecture, vehicles, traps and puzzle markers.

The author describes it as an updated preview from a training run that is still going. In their tests most generations work on the first attempt, especially with English prompts. Full Safetensors weights and the training dataset are planned for after training finishes. The model card also mentions a dataset release on November 3.

## Which file should you download: Q4_K_M or Q8_0?

Both quantizations carry the same updated checkpoint. The model card does not give file sizes or hardware requirements for them, so check the file listing on Hugging Face before you pull anything.

| File | Notes |
|:---|:---|
| Qwen3.5-4B-GLB-3D-Coder-Q4_K_M.gguf | Smaller quantization. |
| Qwen3.5-4B-GLB-3D-Coder-Q8_0.gguf | Higher-precision quantization. |
| mmproj-Qwen3.5-4B-BF16.gguf | Optional vision projector, about 644 MiB. Works with either model file. |

As a general rule, Q8_0 keeps more of the original weights at the cost of memory, and Q4_K_M trades some precision for a smaller footprint. For code that has to run without edits, it is reasonable to try Q8_0 first if your machine can hold it. That is a general heuristic, not something the model card tests.

## Do you need the mmproj vision file?

Only if you want image-reference inputs. For text-only prompts you can skip it. If you do want images, your runtime must support Qwen3.5 multimodal GGUF. Load the projector alongside the model.

The same projector works with Q4_K_M and Q8_0. It is the unmodified Qwen3.5-4B BF16 vision encoder, taken from lmstudio-community/Qwen3.5-4B-GGUF under Apache 2.0, and it was also shipped with CCORE's earlier 3D Coder model.

## What settings and system prompt should you use?

The author's current sampling values are a starting point, not a tuned optimum.

| Parameter | Value |
|:---|:---|
| Temperature | 0.8 |
| Top K | 40 |
| Top P | 0.95 |
| Min P | 0.0 |

Put this in the system message, then send your request as the user message:

```text
Create game assets for the Godot editor, which uses GLB files. Create whatever the user requests as a good-quality low-poly asset using only Three.js in one complete standalone HTML file. The HTML must show the asset and include an Export GLB button that downloads a working GLB. If the requested asset needs animation, add it and include the animation in the exported GLB; otherwise make it static.
```

The prompt asks for low-poly output, a single HTML file, and an Export GLB button. If the asset moves, say so in your request. The system prompt tells the model to include that animation in the exported GLB.

## How do you go from a prompt to a GLB file?

1. Load the GGUF in your runtime and apply the sampling values.
2. Set the system prompt above.
3. Send a short English request, such as `Create a short iron sword with a leather-wrapped grip.`
4. Copy the generated HTML into a file and open it in a browser.
5. Check that the asset renders, then click Export GLB to download the file.

Other prompts from the model card include `Create a stone golem boss with glowing green crystals`, `Create a wooden gate with an opening animation`, and `Build a jungle siege platform`. The author says the published example results came from prompts that were not in the training data.

## How do you import the GLB into Godot?

Godot reads GLB natively, which is why the system prompt names it. Drop the exported `.glb` into your project folder, and the editor imports it as a scene. You can then instance it in a level, attach scripts, or, for animated exports, look for the animation in the imported scene's animation player. The model card shows exported assets after import into the Godot editor, but it does not document a step-by-step import process, so treat the details above as standard Godot behavior rather than something the author specified.

Since the output is generated code, inspect the result in the browser first. A bad export is easier to catch there than inside a game project.

## Why do you need to re-download the weights?

The author replaced both GGUFs with a newer checkpoint on November 2, 2026 at 04:50 UTC, and reports substantially better quality. The filenames stayed the same. If you grabbed a file earlier, a name match tells you nothing. Delete the old weights in LM Studio or whichever app you use, and download them again. The sampling settings and system prompt did not change.

## Is it ready for production use?

Not yet. The author calls it an early checkpoint and lists limits: some generations need a retry or a small code correction, English prompts work better, and behavior may change as training continues. For a game jam prop pipeline or prototyping, that may be acceptable. For anything with strict polygon budgets or consistent art direction, expect to review and edit each asset.

There is a community example of what it can do. Naoki Kishida built a playable 3D RPG demo using assets generated with the model.

## Frequently Asked Questions

### Is this an official Qwen model?

No. It is an independent model by CCORE, fine-tuned from Qwen/Qwen3.5-4B. The base model's vision projector is reused unmodified for image input.

### Does the model generate a GLB file directly?

No. It writes an HTML page with Three.js code that builds the asset. The page includes an Export GLB button that downloads the file. You open the page in a browser to get the GLB.

### Can I use image references?

Yes, if your runtime supports Qwen3.5 multimodal GGUF. Download `mmproj-Qwen3.5-4B-BF16.gguf` (about 644 MiB) and load it with either quantization. Text-only prompts do not need it.

### Why do my results look different from the examples?

Sampling is not deterministic at temperature 0.8, so the same prompt gives different outputs. The model card shows two different results for the same command post prompt. Also confirm you re-downloaded the GGUF after the November 2 update and are using the recommended system prompt.