All stories
Qwen3.5-4B-3D-Coder Ecosystem Community reports

CCORE replaces its tiny 3D coder's GGUF files and flags an LM Studio trap

The independent Three.js fine-tune asks early users to download fresh Q4_K_M or Q8_0 weights and disable MTP in LM Studio.

Sources (3)

Why now An October 26 model-card update replaces both GGUF builds after an author-reported layer issue; this is an actionable preview update, not a completed training release.
CCORE's author-published Three.js scene shows a temple, pink blossom trees, a bridge and a pond.
Author-published rendered scene from the model preview. This is a selected project sample, not an AITrending reproduction or a measurement of one-shot success rates. Source: CCORE

CCORE’s small local 3D coding experiment has a practical update for early downloaders: replace the weights. The independent Qwen3.5-4B-3D-Coder project’s October 26 notice says both its Q4_K_M and Q8_0 GGUF files have been replaced after a layer-related issue. The author says the corrected builds appear to produce more successful one-shot scenes in their latest tests. That is a maintainer report, not an independently reproduced benchmark.

The model fine-tunes Qwen3.5-4B for general coding and interactive Three.js scenes. It is a community project from CCORE, rather than an official Qwen release. Its public repository was created October 25; the subsequent update does not reset that origin date.

There is an important setup detail for LM Studio users: the card says to disable MTP, which can otherwise prevent the model from functioning properly. It suggests temperature between 0.6 and 0.8, top-p between 0.9 and 0.95 and a system prompt that uses the Three.js 0.160.0 core ES module without addon or legacy example imports. These are the author’s workflow recommendations, not a guarantee that generated code will run unchanged.

This remains an early preview with known duplicate-variable bugs. At review time, the card expected training to finish October 28 at 07:30 UTC and promised later FP16 weights, more quantizations, a dataset and a multi-prompt evaluation report. Those are future plans. The immediate news is narrower and useful: refreshed downloadable GGUF builds and a concrete configuration caveat for people trying a four-billion-parameter scene coder locally.