CCORE GLB-3D-Coder - Q3_K_M

Q3_K_M GGUF of CCORE's original u008186 checkpoint, released by its creator on November 5, 2026. This is the GLB asset-coding model, distinct from CCORE's earlier Three.js scene model. Chungulus provides this independent community quantization; it is not an official Qwen or CCORE release.

The original checkpoint has 4,539,265,536 parameters: F32 text weights and BF16 vision weights. We exported its 4,205,751,296-parameter text network to F16 (retaining sensitive F32 tensors), then independently quantized that export to Q3_K_M. The unchanged original mmproj-Qwen3.5-4B-BF16.gguf is included for images. Multi-token prediction is not exported. No low-bit checkpoint was requantized, and no calibration or importance matrix was used.

Use

Use a recent llama.cpp runtime supporting Qwen3.5. The tested build is commit 6414d5b0c9e5895473036c90c2bd305ce4ceccc4. Disable MTP/speculative prediction in frontends that attempt to enable it.

hf download Chungulus/Qwen3.5-4B-GLB-3D-Coder-Q3_K_M-GGUF --local-dir ./Qwen3.5-4B-GLB-3D-Coder-Q3_K_M
llama-completion -m ./Qwen3.5-4B-GLB-3D-Coder-Q3_K_M/Qwen3.5-4B-GLB-3D-Coder-Q3_K_M.gguf -cnv -st --jinja -c 16384 -ngl 99 -n 8192 -p 'Create a low-poly red cube with an Export GLB button in one complete Three.js HTML file.'
llama-mtmd-cli -m ./Qwen3.5-4B-GLB-3D-Coder-Q3_K_M/Qwen3.5-4B-GLB-3D-Coder-Q3_K_M.gguf --mmproj ./Qwen3.5-4B-GLB-3D-Coder-Q3_K_M/mmproj-Qwen3.5-4B-BF16.gguf --image input.png -p 'Describe this image.' --image-min-tokens 1024 -c 8192 -ngl 99

The original model card supplies the creator's asset-generation system prompt and recommends temperature 0.8, top-k 40, top-p 0.95, min-p 0, repeat penalty 1 and thinking enabled. Allow enough output for the complete HTML. Our bounded deterministic checks used temperature 0 and thinking disabled; they do not reproduce the creator's evaluation.

Actual checks and limits

This artifact passed native-runtime arithmetic, JSON output and two image-understanding cases with the included projector: identify the central colored square in opposing red/blue fixtures, using 1024 image tokens, temperature 0 and seed 42. Its generated cube/export example passed JavaScript syntax and expected Three.js/GLTFExporter API checks after optional fence stripping and HTML-entity decoding. Raw and normalized results are recorded separately in validation.json; the raw output is retained locally. example-scene.html is the normalized generated example.

An initial 256-image-token check asked only for the color of the "large square", which can also refer to the white canvas, and did not explicitly fix sampling settings. This format answered red (initial intended-red check passed: True). That result is retained alongside the later, disambiguated two-color checks; unsuccessful initial cases have not been erased. Image understanding is limited evidence, not a guarantee of reliable grounding.

These are short functional checks. Browser rendering, successful GLB download/import, video, full-context behavior and benchmark accuracy were not tested. Inspect and test generated assets in your editor, as the creator recommends. The complete output tensor names/shapes match the unquantized export, the actual GGUF file type was inspected, and source/upload SHA256 digests were verified. quantizatia-manifest.json records the exact source/tool pins and source digests.

License

Apache 2.0. Retain the original license and attribution. Based on Qwen/Qwen3.5-4B and independently fine-tuned by CCORE. The original LICENSE and chat_template.jinja are included.