Apache-2.0llmRuns locally4B (card); 4.54B (HF safetensors F16)

Qwen3.5-4B-3D-Coder: Specs, License & How to Run

Community Apache-2.0 4B Three.js / coding fine-tune. Not official Qwen.

Official Qwen3.5-4B-3D-Coder Sakura Temple Garden Three.js scene screenshot from the ccore card — Official project image. Source · License: Apache-2.0.

License: Apache-2.0
Parameters: 4B (card); 4.54B (HF safetensors F16)
HF likes: 478 (2026-11-02)
First seen: 2026-10-25

TL;DR. Qwen3.5-4B-3D-Coder is a llm model from CCORE (community; not Qwen). Parameters: 4B (card); 4.54B (HF safetensors F16).License: Apache-2.0. It is listed under Claude Code, Cursor, ChatGPT, Closed LLM APIs.

Developer	CCORE (community; not Qwen)
Type	llm
Architecture	Independent fine-tune of Qwen/Qwen3.5-4B (card + cardData.base_model). HF model_type qwen3_5; architecture Qwen3_5ForConditionalGeneration. text_config: 32 layers, hidden_size 2560, 16 Q / 4 KV heads, intermediate 9216, vocab 248320, max_position_embeddings 262144, hybrid linear + full attention (full_attention every 4th layer). Early preview; the card says the author is still training. Distinct from official Qwen/Qwen3.8-27B and from Sperid Labs Idol (image-to-3D splat).
Parameters	4B (card); 4.54B (HF safetensors F16)
Modalities	text, image, code
Languages	en
Context	262,144 tokens (config); llama.cpp example uses 65,536
License	Apache-2.0 — LICENSE on ccore/Qwen3.5-4B-3D-Coder is Apache License 2.0 (cardData.license apache-2.0, opened 2026-11-02). The card states this is an independent CCORE fine-tune of Qwen/Qwen3.5-4B, not an official Qwen release. Hosted Qwen Cloud terms are unrelated. Not legal advice.
Hardware	Not yet verified — Card ships FP16 Safetensors (4 shards), Q4_K_M GGUF, Q8_0 GGUF, and an optional BF16 vision projector (~644 MiB / 676 MB sibling). HF usedStorage is 31,354,574,590 bytes (~29.2 GiB) for the whole repo. Peak VRAM is not independently measured. Card: llama.cpp vision load tested locally with llama.cpp b11889; FP16 text load verified with Transformers 5.20.2. Those are upstream tests, not forkalt runs. (not independently tested)
Hugging Face likes	478 (as of 2026-11-02)
Model card	https://huggingface.co/ccore/Qwen3.5-4B-3D-Coder
First seen	2026-10-25

Qwen3.5-4B-3D-Coder is a community 4B coding preview from CCORE: general-purpose coding with a focus on self-contained Three.js scenes. Official hosts we opened on 2026-11-02: the ungated card ccore/Qwen3.5-4B-3D-Coder (created 2026-10-25T05:43:34Z; last modified 2026-10-30T15:47:46Z; 478 likes / 44,825 downloads), the Apache 2.0 LICENSE, config.json, the named base Qwen/Qwen3.5-4B (1,176 likes / 6,588,471 downloads; Apache-2.0), the MIT dataset ccore/ThreeJS-3D-Scenes (33 likes / 1,781 downloads; 4,146 examples), and the technical report PDF. API gated was false.

The card’s first identification sentence is the one that matters: this is an independent fine-tune of Qwen3.5-4B. It is not an official Qwen release. Official Qwen already on this site: Qwen3.8-27B. Head-to-head facts: Qwen3.5-4B-3D-Coder vs Qwen3.8-27B.

This page hangs off Claude Code alternatives, Cursor alternatives, ChatGPT alternatives, and closed LLM APIs. Those mappings are editorial for a local coding checkpoint. We have no sourced quality comparison against Claude Code, Cursor, or ChatGPT.

It is not Idol (Sperid Labs image-to-3D Gaussian splat) and not SenseNova 6.8 Code2Video (official-ish 27B code-driven video fine-tune of Qwen3.8-27B).

“Early preview,” “still training,” and “one-shot success has improved substantially” stay author / vendor.

What we verified (2026-11-02)

Field	Value	Source
Host	huggingface.co/ccore/Qwen3.5-4B-3D-Coder (not gated)	HF API
Created / last modified	2026-10-25T05:43:34Z / 2026-10-30T15:47:46Z	same
Likes / downloads	478 / 44,825	same
License tag + file	apache-2.0 / Apache License 2.0	HF `cardData` + LICENSE
Relation	finetune of `Qwen/Qwen3.5-4B`	HF `cardData`
Official?	No — card: “not an official Qwen release”	README
Architecture	`qwen3_5` / `Qwen3_5ForConditionalGeneration`	config.json
Params (card / safetensors)	4B / 4,539,265,536 F16	README + HF API
usedStorage	31,354,574,590 bytes (~29.2 GiB)	HF API
text_config	32 layers / hidden 2560 / 16 Q, 4 KV / vocab 248,320	config.json
Context (config)	262,144	`max_position_embeddings`
Files	FP16 ×4 shards, Q4_K_M, Q8_0, optional mmproj	HF siblings
Dataset	ccore/ThreeJS-3D-Scenes · MIT · 4,146 rows · 33 likes	HF dataset API

October 27 GGUF refresh, October 29 FP16 + dataset + PDF, and October 30 optional vision projector are dated on the README we opened. The card says the Q4/Q8 files now contain the October 27 checkpoint.

How to run (from the official card, not tested)

Copied names only. forkalt has not loaded a checkpoint and has not rendered a Three.js file.

GGUF (card): temperature 0.6–0.8, top_p 0.90–0.95, top_k 20–40, repeat penalty 1.0. LM Studio: disable multi-token prediction (MTP); the card says LM Studio enables it by default.

FP16 Transformers (card): temperature 0.6, top_k 20, top_p 0.95, repetition penalty 1.0; author says loading + text generation was verified with Transformers 5.20.2.

Vision GGUF (optional, 2026-10-30): download mmproj-Qwen3.5-4B-BF16.gguf (~644 MiB; unmodified projector from lmstudio-community/Qwen3.5-4B-GGUF, Apache-2.0) and load it with llama.cpp:

llama-server \
  -m Qwen3.5-4B-3D-Coder-Q4_K_M.gguf \
  --mmproj mmproj-Qwen3.5-4B-BF16.gguf \
  --n-gpu-layers all --flash-attn on \
  --ctx-size 65536 --parallel 1

Card: tested with llama.cpp b11889. Image tokens share the context window. That is an upstream short test, not a forkalt run.

Three.js system prompt copied from the card: use only the Three.js core ES module from cdnjs 0.160.0; do not import addons or legacy examples/js paths.

License

LICENSE is Apache License 2.0. Card tag apache-2.0. The dataset is MIT. The optional projector is Apache-2.0 from the lmstudio-community Qwen3.5-4B GGUF. Not legal advice.

What we did not invent

Scene quality, “one-shot success,” and known-issue notes (duplicate variable names) stay author.
We do not rank this card against Claude Code, Cursor, Qwen3.8-27B, Spark-X2.6-4B, or Flint.
Peak VRAM is not independently measured.
This is not official Qwen and not an image-to-3D splat.

Related: Qwen3.5-4B-3D-Coder vs Qwen3.8-27B — Qwen3.8-27B (official 27B VLM) — Spark-X2.6-4B (~4B on-device coding/agent) — Flint (39.8B/2.5B coding MoE) — SenseNova 6.8 Code2Video (code-driven video; different job) — Idol (image-to-3D splat; different job). Serving: vLLM. Agents: Cline, OpenCursor. Catalog: home. Corrections: contact.

Deploy options

Comparisons

Qwen3.5-4B-3D-Coder vs Qwen3.8-27B

FAQ

Is this an official Qwen model?

No. The card says it is an independent fine-tune by CCORE of Qwen/Qwen3.5-4B and “is not an official Qwen release.” Official Qwen cards already on this site include Qwen3.8-27B. Compare Qwen3.5-4B-3D-Coder vs Qwen3.8-27B.

Are the weights public, and what is the license?

Yes. huggingface.co/ccore/Qwen3.5-4B-3D-Coder is ungated (created 2026-10-25T05:43:34Z; last modified 2026-10-30T15:47:46Z; 478 likes / 44,825 downloads in the 2026-11-02 API check). LICENSE is Apache License 2.0. cardData.license is apache-2.0. The training set ccore/ThreeJS-3D-Scenes is MIT (33 likes). Not legal advice.

How big is it?

The card calls it a 4B preview. Hugging Face safetensors.parameters total 4,539,265,536 F16 (~4.54B). The official base Qwen/Qwen3.5-4B reports 4,659,865,088 BF16 tensors. We copy both the 4B marketing line and the tensor sums; we do not pick a winner.

Does it replace Claude Code, Cursor, or Qwen3.8-27B?

No sourced bake-off. This page identifies a community coding / Three.js fine-tune. “One-shot success has improved substantially” is the author’s qualitative note, not a leaderboard. See Claude Code alternatives and Qwen3.5-4B-3D-Coder vs Qwen3.8-27B.

Sources & verification

Last verified 2026-11-02 · Last updated 2026-11-02· Edited by forkalt editors · Editorial policy

License notes are not legal advice. Read the original LICENSE. Trademark & disclaimer.