Qwen3.5-4B-3D-Coder - Q6_K

Q6_K GGUF of CCORE's original FP16 checkpoint, quantized independently from its full-precision export. The upstream checkpoint has 4,539,265,536 parameters including vision; this text GGUF has 4,205,751,296 parameters. The unchanged upstream BF16 vision projector is included. This is a community quantization of CCORE's fine-tune, not an official Qwen or CCORE release.

Use

Use a recent llama.cpp build supporting Qwen3.5. Tested with commit 6414d5b0c9e5895473036c90c2bd305ce4ceccc4. Multi-token prediction is not exported; disable MTP/speculative prediction in frontends that try to enable it.

hf download Chungulus/Qwen3.5-4B-3D-Coder-Q6_K-GGUF --local-dir ./Qwen3.5-4B-3D-Coder-Q6_K
llama-completion -m ./Qwen3.5-4B-3D-Coder-Q6_K/Qwen3.5-4B-3D-Coder-Q6_K.gguf -cnv -st --jinja -c 8192 -ngl 99 -p 'Create a minimal Three.js scene with a rotating red cube.'
llama-mtmd-cli -m ./Qwen3.5-4B-3D-Coder-Q6_K/Qwen3.5-4B-3D-Coder-Q6_K.gguf --mmproj ./Qwen3.5-4B-3D-Coder-Q6_K/mmproj-Qwen3.5-4B-BF16.gguf --image input.png -p 'Describe this image.' -c 8192 -ngl 99

Checks and provenance

Actual native-runtime tests covered arithmetic, JSON output and image understanding using the included projector. A generated Three.js sample was checked for JavaScript syntax after removing markdown fences and decoding HTML entities. Some raw outputs use escaped HTML; raw and normalized syntax results are reported separately. Browser rendering, video, full-context behavior and benchmark accuracy were not evaluated.

See validation.json, example-scene.html and quantizatia-manifest.json for the checks, example and source/tool revisions. No calibration dataset, importance matrix or requantization was used. The original source and projector hashes were verified before conversion. Apache 2.0 terms are included in LICENSE.