Independent author CCORE's preview programming model trained on Qwen3.5-4B: Describe a game asset (character, weapon, prop, building module, vehicle, etc.) in English, and it generates a single-file Three.js HTML with an "Export GLB" button. Open it in a browser to view the model and download the .glb file, which can then be imported into editors like Godot; assets requiring animation will have their animations exported as well. This is one of two models by the same author, the other being Qwen3.5-4B-3D-Coder (which generates entire 3D scenes) featured on 10 29. This repository was created on 11 1 and was not included in previous issues.
Score increase 79 · Likes 79 · Downloads 381
Reasons
- The goal shifts from "viewable scenes" to "reusable assets": The author states that the training data covers 25 types of game assets, with example prompts provided for each type in the model card. The gallery samples use prompts outside the training set and show the results after importing into Godot.
- On 11 2, the author released a new checkpoint, claiming a significant improvement in generation quality, with most outputs being usable on the first try. On the same day, a vision projection file was added, allowing the use of reference images. Additionally, a community developer used assets generated by this model to create a playable web-based 3D RPG demo (kishida.github.io/webdemos/rpg-3d-glb).
- Small size: Q4_K_M is 2.71 GB and Q8_0 is 4.48 GB, allowing it to run locally using LM Studio or llama.cpp.
Usage
- Load Q4_K_M or Q8_0 GGUF in LM Studio or llama.cpp. Use the system prompt exactly as provided in the model card (requiring the use of Three.js for low-poly assets, a single HTML file, and an Export GLB button). Write a one-sentence English requirement in the user message, for example: Create a short iron sword with a leather-wrapped grip.
- Sampling parameters: temperature 0.8, top_k 40, top_p 0.95, min_p 0. Specify animation requirements in the prompt if needed.
- To use reference images, download mmproj-Qwen3.5-4B-BF16.gguf (approx. 644 MiB, the original Qwen3.5-4B projector, unmodified), which is compatible with both Q4 and Q8. If you downloaded the GGUF before 11 2 04:50 (UTC), delete it and re-download it; the filename has not changed.
Risks
- Early preview, training is ongoing, and there are no benchmark scores. The author notes that some generations may require retries or manual code adjustments, and English prompts perform better than other languages. Full safetensors weights and the training dataset will be released only after training is complete.
- Image understanding capabilities come from the original Qwen3.5-4B projector, with no specific visual training for assets; the output is in a low-poly style.
- This is a personal fine-tune, not an official Qwen version. The generated HTML is code written by the model; review it before opening.