{"slug": "qwen3-5-4b-glb-3d-coder-run-this-local-model-for-game-assets", "title": "Qwen3.5-4B-GLB-3D-Coder: Run This Local Model for Game Assets", "summary": "Independent developer ccore released Qwen3.5-4B-GLB-3D-Coder, a 4B-parameter fine-tune of Qwen's Qwen3.5-4B that generates standalone Three.js HTML/JavaScript scenes with a built-in Export GLB button, producing assets intended for import into Godot. The model ships as two GGUF quantizations, Q4_K_M and Q8_0, for local inference via LM Studio or llama.cpp, and a November 2, 2026 checkpoint update replaced both GGUF files with higher visual quality using the same sampling settings and system prompt. Training covers 25 game-asset domains, and an optional BF16 vision projector (644 MiB) adds image-reference input for runtimes supporting Qwen3.5 multimodal GGUF.", "body_md": "# Qwen3.5-4B-GLB-3D-Coder: Run This Local Model for Game Assets\n\nA guide to Qwen3.5-4B-GLB-3D-Coder, a local GGUF model that writes Three.js scenes and exports GLB assets for Godot game projects.\n\n## What is Qwen3.5-4B-GLB-3D-Coder?\n\nQwen3.5-4B-GLB-3D-Coder is a 4B parameter fine-tune of Qwen’s Qwen3.5-4B model, trained to generate standalone Three.js HTML/JavaScript scenes that export directly to GLB files. The GLB output is meant to drop straight into game engines like Godot, which makes it a narrow but practical tool for solo developers and small teams who need low-poly 3D assets without touching a modeling program. It’s released by an independent developer under the handle ccore, not by the Qwen team, and is distributed as GGUF files for local inference.\n\n## TL;DR\n\n- **Qwen3.5-4B-GLB-3D-Coder** is a 4B parameter Qwen3.5 fine-tune that writes self-contained Three.js HTML files with a built-in Export GLB button, producing assets ready for Godot.\n- The model ships as **two GGUF quantizations** , Q4_K_M and Q8_0, so it runs on consumer GPUs or even CPUs through tools like LM Studio.\n- A **November 2, 2026 checkpoint update** replaced both GGUF files with noticeably better visual quality, using the same sampling settings and system prompt as before.\n- The creator recommends specific **generation settings** (temperature 0.8, top K 40, top P 0.95, min P 0.0) and a fixed system prompt that tells the model it’s building Godot-compatible GLB assets.\n- Training covers **25 game-asset domains** , from characters and bosses to traps, watercraft, and festival props, each with example prompts in the model card.\n- An **optional vision projector** (BF16, 644 MiB) adds image-reference input for runtimes that support Qwen3.5 multimodal GGUF, but it’s not required for text-only use.\n- The creator plans a much larger follow-up: a 100,000-example dataset with context lengths from 8K to 128K tokens, aimed at reliable one-shot generation and eventual use inside an agent harness.\n\n*regulatory compliance and security.*\n\n- Name\n- Maria Alvarez\n- Date of birth\n- SSN\n- Diagnosis\n- Balance due\n- $412.80\n\n## How does the model generate 3D assets?\n\nThe model doesn’t output a 3D file directly. Instead, it writes a complete HTML document containing a Three.js scene: the geometry, materials, and (when requested) animations for the asset you describe. You save that HTML, open it in a browser, and the page includes an Export GLB button that packages the scene into a GLB file. That GLB is what you import into Godot or another engine.\n\nThis two-step process (generate HTML, then export GLB in-browser) is a deliberate design choice. It lets the model focus on writing valid Three.js code rather than producing a binary format directly, and it gives you a visual preview before committing to the export. According to the model card, most generations work on the first attempt, particularly with prompts written in English, though some asset requests need a retry or a small manual code fix.\n\n## What do you need to run it locally?\n\nThe model is distributed as GGUF files, which are built for local inference through tools like LM Studio, llama.cpp, or similar runtimes. Two quantizations are available:\n\n- **Q4_K_M** : a smaller, more compressed version that trades some precision for lower memory use.\n- **Q8_0** : a higher-precision version that needs more VRAM or RAM but preserves more of the original model’s behavior.\n\nBecause the base model is only 4B parameters, both quantizations are small enough to run on a single consumer GPU, and Q4_K_M can likely run on more modest hardware or CPU-only setups. If you want to feed the model image references instead of just text prompts, you can load the separate BF16 vision projector (644 MiB) alongside either quantized model, provided your runtime supports Qwen3.5’s multimodal GGUF format. That projector is the original, unmodified Qwen3.5-4B vision encoder released under Apache 2.0, not something custom-built for this fine-tune.\n\n## How do you set it up in LM Studio?\n\nOnce you’ve downloaded one of the GGUF files into LM Studio (or pointed your runtime at it), the setup has two parts: the system prompt and the sampling settings.\n\nThe system prompt instructs the model to build Godot-compatible, low-poly assets using only Three.js in a single HTML file, to include a working Export GLB button, and to add animation only when the user asks for it. This prompt is fixed and documented in the model card, so it’s worth copying exactly rather than paraphrasing.\n\nFor sampling, the creator’s tested settings are:\n\n| Parameter | Value | \n|---|---|\n| Temperature | 0.8 | \n| Top K | 40 | \n| Top P | 0.95 | \n| Min P | 0.0 | \n\nThese aren’t hard requirements. The model card notes they’re a starting point and invites experimentation. After setup, you send a short, specific English request, such as “Create a swamp goblin scout with a wooden shield,” and the model returns the HTML file.\n\n## What kinds of assets can it make?\n\nAuditable. Scalable.\n\nThe current checkpoint was trained across 25 distinct game-asset categories, which gives a sense of how broad the intended use case is. These span characters and NPCs, enemies, bosses, companions, weapons and armor, items and pickups, props and furniture, nature elements, modular architecture pieces, interactive objects, ground/air/water transport, industrial machinery, lab and medical props, urban fixtures, sports and playground equipment, farming assets, market and cuisine props, museum displays, traps and hazards, festival decorations, battlefield scenery, underwater structures, and puzzle or gameplay markers.\n\nExample outputs shown in the model card include a jungle siege platform, a weathered snow siege platform, a blue goblin brute, a red orc leader, and a reinforced mountain command post generated twice from the same prompt to show variation. Several of these were also shown successfully imported into the Godot editor after GLB export, which is the real test of whether the pipeline holds up end to end. A community member, Naoki Kishida, used assets from the model to build a playable 3D RPG demo, which is a reasonable signal that the output is usable in an actual shipped project rather than just a demo screenshot.\n\n## Is this model worth using right now?\n\nFor prototyping and asset placeholders, yes, with caveats. It’s free, runs locally, and the GLB-to-Godot workflow is straightforward once you’ve got the system prompt and settings dialed in. The fact that it’s a 4B model means it’s fast and cheap to run compared to larger code or multimodal models, and the creator reports that most first-attempt generations succeed.\n\nThe caveats: this is explicitly labeled a “preview” and an independent, unofficial release, not something from the Qwen team. The dataset so far has focused on simple prompts, which the creator says makes it accessible but also limits how well it generalizes to highly detailed or unusual requests. The creator has already flagged this limitation and is doubling the training dataset, with a much larger 100,000-example dataset planned for a future version aimed at more reliable one-shot generation and eventual integration into an agent workflow. If you need production-quality, highly customized assets today, expect to do manual cleanup in Godot after import. If you need quick low-poly placeholders or a fast way to populate a prototype, it’s a reasonable tool to add to a local pipeline.\n\n## Frequently Asked Questions\n\n### What base model is Qwen3.5-4B-GLB-3D-Coder built on?\n\nIt’s a fine-tune of Qwen/Qwen3.5-4B, an open 4B parameter model released by the Qwen team. The GLB-3D-Coder variant is an independent fine-tune by a developer going by ccore, not an official Qwen release.\n\n### Does it output GLB files directly?\n\nNo. It generates a standalone HTML file containing a Three.js scene. You open that file in a browser and click the Export GLB button it includes to produce the actual GLB file for import into Godot or another engine.\n\n### What hardware do I need to run it?\n\nBecause it’s a 4B parameter model distributed in GGUF format (Q4_K_M and Q8_0), it’s designed for local inference on consumer hardware through tools like LM Studio. Q4_K_M needs less memory than Q8_0, making it the better choice for lower-end GPUs or CPU-only setups.\n\n### Can I use image references instead of text prompts?\n\nYes, if your runtime supports Qwen3.5’s multimodal GGUF format. You load the separate BF16 vision projector file (644 MiB) alongside the Q4_K_M or Q8_0 model. It’s optional and not needed for standard text-to-asset generation.\n\n### Is the training dataset available?\n\nThe creator has stated plans to release the complete dataset, including additional examples from an in-progress expansion, alongside full Safetensors weights, following the same approach taken with a previous 3D Coder model release.", "url": "https://wpnews.pro/news/qwen3-5-4b-glb-3d-coder-run-this-local-model-for-game-assets", "canonical_source": "https://www.mindstudio.ai/blog/qwen3-5-4b-glb-3d-coder-local/", "published_at": "2026-11-04 00:00:00+00:00", "updated_at": "2026-11-04 10:27:07.664318+00:00", "lang": "en", "topics": ["ai-tools", "generative-ai", "large-language-models", "ai-products"], "entities": ["Qwen3.5-4B-GLB-3D-Coder", "Qwen", "ccore", "Godot", "Three.js", "LM Studio", "llama.cpp"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/qwen3-5-4b-glb-3d-coder-run-this-local-model-for-game-assets", "markdown": "https://wpnews.pro/news/qwen3-5-4b-glb-3d-coder-run-this-local-model-for-game-assets.md", "text": "https://wpnews.pro/news/qwen3-5-4b-glb-3d-coder-run-this-local-model-for-game-assets.txt", "jsonld": "https://wpnews.pro/news/qwen3-5-4b-glb-3d-coder-run-this-local-model-for-game-assets.jsonld"}}