A specialized 4B model for text-to-GLB game assets goes head to head with DeepSeek V4.1 Flash on identical 3D asset prompts.
What is CCORE GLB-3D-Coder, and how does it compare to DeepSeek V4.1 Flash for making 3D assets? #
CCORE GLB-3D-Coder is a 4B parameter coding model built specifically to turn text prompts into exportable GLB game assets, while DeepSeek V4.1 Flash is a general-purpose reasoning model that can also write Three.js code when asked. In a direct comparison using identical prompts, both models produced usable low-poly 3D scenes, but GLB-3D-Coder was purpose-trained on a dataset of rendered, validated GLB examples, giving it an edge on mechanical plausibility (cable routing, joint placement, part mounting) that a general model has to infer from scratch each time.
TL;DR #
- CCORE GLB-3D-Coder is a 4B model fine-tuned from Qwen3.5-4B that generates a complete Three.js HTML page with a built-in Export GLB button, producing assets ready to drop into Godot or any GLB-compatible editor.
- The comparison pitted it against DeepSeek V4.1 Flash running at high reasoning effort, using the exact same prompts for both models, from short one-liners to detailed multi-paragraph specs.
- The model was trained on CCORE ThreeJS-3D-GLB-Assets , a released dataset of 3,918 tested and rendered examples spanning 40 domains, which is likely why its outputs tend to hold together mechanically (cables, mounts, rail spacing) better than a general model’s first pass.
- Generation is token-heavy : training examples reached up to 72K tokens, average responses run around 20K tokens, and the model card explicitly warns against capping output at 8K tokens or disabling extended thinking.
- The weights ship in Q4_K_M, Q8_0, and original Safetensors formats under Apache 2.0, with the Q8_0 quantization used for the published evaluation.
- A playable browser game, RIFTWARD: Guardians of the Ancient , was built using assets generated by this model, showing the pipeline works beyond isolated test prompts.
- This is a solo, self-funded open release, with the dataset, model weights, and a roadmap toward a much larger 100,000-example training set published alongside it.
Nothing gets lost.
How does GLB-3D-Coder actually generate a 3D asset? #
The model doesn’t output a 3D file directly. Instead, given a text prompt, it writes a full standalone HTML page that embeds Three.js, builds the requested object or scene programmatically (meshes, materials, positioning, optional animation), renders a live preview in the browser, and exposes an “Export GLB” button. Clicking that button triggers a GLB file download, which can then be imported into Godot or another engine that reads the GLB format.
This matters because it means the model is really writing code, not predicting pixels or vertices directly. Three.js scene construction is explicit and structured, which plays to the strengths of a model fine-tuned on code. The system prompt used for the model instructs it to build “good-quality low-poly” assets as Three.js-only HTML, add animation only when requested, and always include a working export path.
Because the model is trained on validated, rendered examples rather than raw scraped code, it has seen complete, working patterns for 40 different asset domains: characters, weapons, vehicles, modular architecture, props, environmental scenery, and more. That domain coverage shows up in the list of suggested prompts, which ranges from a village blacksmith to a cargo airship to a hidden spiked floor trap.
What did the side-by-side comparison actually show? #
The comparison used identical prompts sent to both models, with DeepSeek V4.1 Flash set to high reasoning effort and GLB-3D-Coder run at its default settings. Several prompts were tested, from short single-sentence requests to long, detailed briefs specifying materials, part placement, and mechanical constraints.
One example, “Create a coral-armored lighthouse keeper,” tested how each model handles a short, imaginative prompt with no explicit structure. Another, a detailed brief for a “magnetic-rail rescue engineer,” specified an adult figure in a charcoal work suit with orange safety panels, insulated gloves, a coil-testing wand, a folded rail clamp, and a power pack connected to a tool belt by a visible cable, explicitly asking for mechanically plausible proportions and cable routing.
A third prompt asked for a compact bioluminescent cave tram station, including a raised platform, a two-seat tram on rails, a hand-operated winch, and six caged crystal lamps, again with an explicit ask for coherent wheel flanges, rail spacing, and lamp mounting.
These detailed, constraint-heavy prompts are a useful stress test because they require the model to reason about physical plausibility, not just aesthetics. A general-purpose model like DeepSeek can write correct Three.js syntax but still has to invent sensible geometry relationships (how a cable attaches to a belt, how a wheel flange sits on a rail) from general knowledge. A model trained specifically on rendered, tested 3D asset examples has effectively seen thousands of these part-to-part relationships already resolved.
Is a specialized 4B model actually worth it over a larger general model? #
For this narrow task, the case for a specialized model rests on three things: reliability of the export pipeline, consistency of mechanical detail, and resource cost. Without shadow IT.
Every app, approved
✓
✓
✓
✓
✓
On resource cost, the gap is clear. A 4B model fits comfortably in local GPU memory, especially at Q4_K_M or Q8_0 quantization, while DeepSeek V4.1 Flash is a larger model typically run through an API or hosted endpoint. If the workflow is “iterate on dozens of game assets,” running a small local model repeatedly is cheaper and faster to loop on than hitting a hosted reasoning model each time, even a fast one.
On reliability, the specialized model was trained end-to-end on the specific output format it needs to produce: Three.js HTML with a working Export GLB button and animation baked into the export when requested. That’s a narrower, more controllable target than asking a general model to improvise an export mechanism correctly every time. The model card itself flags that some requests still need a retry or small code correction, and recommends inspecting the exported GLB in the target editor because materials, lighting, and animation can look different from the in-browser preview. That’s an honest limitation, not a guarantee of perfection, but it’s a known failure mode with a known fix (retry, inspect, adjust).
Where a general model like DeepSeek V4.1 Flash still has an edge is flexibility. It can shift between tasks, write more conventional application code, reason about more varied prompt formats, and isn’t limited to the GLB/Three.js niche. If the job is strictly “generate Godot-ready assets in bulk,” the specialized model is the more purpose-fit tool. If the job varies, a general model’s breadth matters more than narrow consistency.
What should you know before running GLB-3D-Coder yourself? #
The model needs generous output limits. Training examples ran as long as 72K tokens, and typical responses average around 20K tokens. An 8K token output cap, common in many default runtime configs, will cut the model off mid-thought or mid-HTML, producing broken or incomplete code. The recommended setting is unlimited output (-1 where supported), with extended thinking left enabled and unrestricted. The model hasn’t been validated with thinking disabled or constrained, so that’s not a safe shortcut for faster generation.
Recommended sampling settings are a temperature of 0.8, top-K of 40, top-P of 0.95, min-P of 0.0, and a repeat penalty of 1.0, matching the settings used in the published evaluation. The weights are available as Q4_K_M and Q8_0 GGUF quantizations for local inference, plus the original Safetensors checkpoint (five shards, roughly 17.49 GB, FP32 text weights with BF16 vision weights) for anyone wanting the unconverted model. An optional BF16 vision projector is available separately for multimodal image-input use cases in compatible runtimes.
Anyone who downloaded the model before its latest update should redownload it, since the checkpoint was revised without a filename change.
Frequently Asked Questions #
What file format does GLB-3D-Coder actually output?
It outputs a complete standalone HTML file containing Three.js code. Opening that HTML in a browser shows a live preview of the asset and an Export GLB button, which downloads the actual GLB file you import into Godot or another 3D editor.
Does GLB-3D-Coder support animated assets?
Yes, if you ask for it explicitly. The system prompt instructs the model to add animation when requested and bake it into the exported GLB; otherwise it produces a static asset. It’s worth double-checking the exported file in your editor to confirm the animation carried over correctly.
Why does GLB-3D-Coder need such long output limits?
demandsof enterprises.
Passed
*✓*SOC 2 Type 2
*✓*HIPAA
*✓*CAIQ attestation
*✓*OCSF logging
*✓*Role-based access
*✓*Self-hosted
The model was trained on examples up to 72,000 tokens long, and typical generations average around 20,000 tokens because they include both reasoning and a full HTML/Three.js file. Capping output at a common default like 8,000 tokens can truncate the response before the code is complete.
Is DeepSeek V4.1 Flash bad at 3D asset generation?
No, it produced usable Three.js output in the comparison when run at high reasoning effort. The difference is that it’s a general-purpose model improvising a GLB export workflow, while GLB-3D-Coder was trained specifically on thousands of rendered, validated examples of that exact workflow, which shows up in details like cable routing and mechanical part placement.
Is the training dataset for GLB-3D-Coder public?
Yes. The CCORE ThreeJS-3D-GLB-Assets dataset, containing 3,918 tested and rendered examples across 40 domains with validated GLB exports, was released alongside the model under an open license.