How to Run Qwen3.5-4B-3D-Coder Locally: The Beginner's Guide

CodexFuse Team · 2026-11-03T13:36:49.072Z · 14 min read min read

Run Qwen3.5-4B-3D-Coder on your own machine: the Ollama one-liner, LM Studio setup with MTP off, the pinned Three.js system prompt, and how to fix the duplicate-variable bug.

If you want to run Qwen3.5-4B-3D-Coder locally, you are asking a very specific 4-billion-parameter model to do one very specific job: write complete Three.js scenes as single HTML files, on your own machine, from plain-English prompts. It is an independent fine-tune of Qwen3.5-4B by solo developer Caio Giovannetti, publishing on Hugging Face as CCORE, trained on 4,146 scene examples and released as Apache-2.0 open weights on October 25, 2026. The recommended download is just 2.71 GB, so an ordinary laptop can run it.

Two things to know before you start: the model is still an early preview — CCORE writes "I am still training" on the model card, and the current checkpoint is dated October 27, 2026 — and its one known bug is small but predictable: it sometimes declares the same variable name twice, which breaks the generated code. This qwen3.5-4b-3d-coder tutorial walks you from zero to your first rendered scene on your own hardware, including how to fix that bug when it appears. If you are new to local weights entirely, read the Mistral Large 4 run guide first for the general idea of running a model locally; the 3D-specific steps below are what make this model different.

What Is Qwen3.5-4B-3D-Coder, Exactly?

Qwen3.5-4B-3D-Coder is a 4-billion-parameter language model fine-tuned to generate Three.js scene code. You describe a scene in words — "a cozy low-poly campsite at dusk with a campfire" — and it returns a complete HTML file you open in a browser and see the 3D scene render. It was trained on 4,146 scene examples (3,933 for training, 213 for validation), including prompts, thinking traces, and the resulting HTML and JavaScript, published as the MIT-licensed dataset ccore/ThreeJS-3D-Scenes. The model weights themselves are Apache-2.0, which means you can use, modify, and ship them commercially.

CCORE is explicit that this is not an official Qwen release — the CCORE 3D Coder is an independent fine-tune by one developer whose Hugging Face collection is called "Small Models for 3D Creation" and also includes a sibling text-to-GLB game-asset model.

Which Setup Fits Your Machine?

There are three verified ways to run this model, and the right one depends on how you like to work. All three read the same GGUF weight files; they differ only in how much control they give you. Pick one row and follow its path — you do not need all three.

The rest of this guide uses the Ollama path because it is the shortest — one install, one command, one prompt — but the settings and system prompt below apply no matter which runtime you choose.

How Do You Run Qwen3.5-4B-3D-Coder Locally?

You run Qwen3.5-4B-3D-Coder locally in five steps: install Ollama, pull the model with a single command, set the Three.js system prompt the model card pins, send a generation prompt with the card's recommended settings, and save the returned HTML to a file you open in a browser. Here is the full arc, beginner to rendered scene.

1. Install Ollama. Download the installer from ollama.com for your operating system. Linux users can use the official install script; Windows and macOS have standard installers. Verify with ollama --version in a terminal.
2. Pull and run the model. The fastest route is the direct Hugging Face reference: ollama run hf.co/ccore/Qwen3.5-4B-3D-Coder:Q4_K_M. A community packaging also exists at ollama run maternion/qwen3.5-3d-coder:4b, which wraps the same Q4_K_M weights with the card's system prompt baked in.
3. Set the system prompt. The model expects a strict system prompt that pins the Three.js version — paste it exactly (see the next section).
4. Ask for a scene. Describe what you want in one concrete paragraph: setting, objects, lighting, mood, and any animation.
5. Save and open the HTML. Copy the model's output into a file ending in .html and open it in any modern browser. The scene needs an internet connection, because the generated page loads Three.js from a CDN.

Step 1: Install Ollama

Ollama is the runtime that downloads, quantizes, and serves the model on your machine. If you have never run a local model before, the self-hosted AI chatbot guide walks through the same Ollama install plus a chat interface; the steps here are the model-specific part that guide does not cover. Install from ollama.com, then open a terminal and confirm the command works before you pull a 2.71 GB file.

Step 2: Pull the Model With the Right Command

The model card publishes two GGUF quantizations: Q4_K_M at 2,708,804,096 bytes (2.71 GB) and Q8_0 at 4,482,402,816 bytes (4.48 GB). Start with Q4_K_M — it is the quantization the community has tested most, and on a laptop the difference in output quality is smaller than the difference in download time and RAM. The qwen3.5-4b-3d-coder Ollama route below is the fastest way to get either file running.

The direct pull command, verified by independent tester webdiga on October 27, 2026, is:

ollama run hf.co/ccore/Qwen3.5-4B-3D-Coder:Q4_K_M

That hf.co prefix tells Ollama to resolve the file straight from the Hugging Face repo, so you always get the current checkpoint. Alternatively, the community packaging ollama run maternion/qwen3.5-3d-coder:4b offers the same weights plus a 4b-q8_0 tag for the near-lossless 4.5 GB version. One caution from fresh research: a second page under the username "maternion" mirrors the first with identical content, and authorship beyond the username is unverifiable — treat the maternion page, the one indexed in search results, as the reference.

Step 3: Use the Pinned System Prompt

This is the step most beginners skip, and it is the one that matters most. The model was trained to emit scenes against one specific Three.js build, and the card pins it in a system prompt you should use verbatim:

Use only the Three.js core ES module from https://cdnjs.cloudflare.com/ajax/libs/three.js/0.160.0/three.module.min.js. Do not import addons or legacy examples/js paths; implement any camera controls directly inside the HTML.

Why it matters: if the model reaches for a newer Three.js version or an addons path it was not trained on, the scene breaks in ways that look like your fault. Every qwen3.5-4b-3d-coder Three.js scene the model emits assumes this exact 0.160.0 build, so pin it before your first prompt. The community Ollama packaging bakes this prompt in; on the direct hf.co route, set it as the system message yourself.

FROM hf.co/ccore/Qwen3.5-4B-3D-Coder:Q4_K_M
PARAMETER temperature 0.7
PARAMETER top_p 0.95
PARAMETER top_k 20
PARAMETER repeat_penalty 1.0
PARAMETER num_ctx 16384
SYSTEM "Use only the Three.js core ES module from https://cdnjs.cloudflare.com/ajax/libs/three.js/0.160.0/three.module.min.js. Do not import addons or legacy examples/js paths; implement any camera controls directly inside the HTML."

Step 4: Ask for Your First Scene

Write the prompt the way the training data was written: one concrete paragraph, setting first, then objects, then lighting and mood, then motion. A good first prompt looks like this:

A cozy low-poly campsite at dusk. A canvas tent, a stone campfire ring with animated flames, two log benches, pine trees around the edges, and fireflies drifting in the air. Warm orange firelight against a deep blue evening sky. The camera slowly orbits the campsite.

Keep temperature in the card's 0.6–0.8 band for first attempts — CCORE notes that higher temperatures make the duplicate-variable bug more likely. Expect to wait: the independent CPU-only test run took 18 minutes 25 seconds to generate 10,609 tokens. This is a "start it and make coffee" model on CPU, not an instant one — but it is the first practical three.js scene generation local llm workflow that works end to end on ordinary hardware.

Step 5: Save the HTML and Open It in a Browser

The model's output is a complete HTML document. Save it with a .html extension and open it in Chrome, Edge, Firefox, or Safari. Two things to check: the browser needs internet access, because the page loads Three.js from the cdnjs CDN link in the system prompt — an offline browser will show a blank page. And if nothing renders, open the browser console (F12) before assuming the model failed; the most common errors are one-line JavaScript fixes covered in the debugging section below.

What Settings Should You Use for Qwen3.5-4B-3D-Coder?

The model card publishes exact generation settings, and the community has confirmed a working set on top of them. Use the card's ranges as your starting point and the tested picks when you want a known-good configuration.

The card's settings apply to both quantizations identically. If a generation stops mid-file, the usual cause is context, not the model failing — bump num_ctx and retry before changing anything else.

How Do You Fix Duplicate-Variable Errors and Other Bugs?

The model card's entire "Known issues" section is one honest line: the bugs observed so far have generally involved duplicate variable names, not major structural problems. CCORE puts a number on it in the repo's discussion thread: around a 60% one-shot success rate on complex scenes so far, with failures that are "sometimes unused variables or the same variable declared twice," especially at higher temperatures. For context, he adds that the base 4B model "struggled to even generate a decent chair" in his tests — the fine-tune is a large step up, but code review is still part of the workflow.

When a scene fails, the independent test run documented exactly what to look for:

1. Identifier 'X' has already been declared — the model wrote the same block twice. Delete the duplicate declaration; everything else usually works.
2. Reading .material on something already a material — the model confused a mesh with its material. Remove the extra .material hop.
3. The scene gets slower over time — the model created new marker meshes inside the animation loop every frame. Move the creation outside the loop and only update positions inside it.

Three habits keep you out of trouble: keep temperature at 0.7 or below, re-download the weights if your copy predates the October 27 checkpoint (an October 26 "fix layer issue" update changed the GGUFs — the card says to re-download earlier copies), and treat every output as a draft that gets one quick console check. This is preview software; the review step is the price of admission, not a sign you did something wrong.

How to Run Qwen3.5-4B-3D-Coder in LM Studio

The qwen3.5-4b-3d-coder LM Studio route is the graphical alternative to the command line: if you prefer a graphical interface, LM Studio runs the same GGUF files with point-and-click loading — but there is one card-documented gotcha you must handle first. Disable multi-token prediction (MTP) for this model. LM Studio enables MTP by default, and the model card explicitly warns that the model will not work properly with it on. New to the app? The LM Studio beginner guide covers installing it and loading any GGUF; the steps below are the 3D-Coder-specific additions.

1. In LM Studio, search the model browser for ccore/Qwen3.5-4B-3D-Coder and download the Q4_K_M quantization (2.71 GB).
2. Before loading, open the model settings and turn off multi-token prediction.
3. Set the system prompt to the pinned Three.js prompt from Step 3 above.
4. Set temperature 0.7, top-p 0.95, top-k 20 — the card's ranges — and send your scene prompt.
5. Copy the returned HTML into a .html file and open it in a browser with internet access for the Three.js CDN.

The duplicate-variable debugging rules from the previous section apply identically here — the bug is in the model's weights, not your runtime, so it appears on every path.

What Can't Qwen3.5-4B-3D-Coder Do Yet?

Honest limits, all from the card or the author's own words. It is an early preview and CCORE is still training it — three checkpoints shipped in three days (October 25, 26, and 27), so pin or re-download rather than trusting a stale copy. Complex scenes land first-try about 60% of the time by the author's own count, and the rest need the one-line fixes above. The generated pages are not offline: they pull Three.js from a CDN, so a sandboxed or offline browser shows nothing.

It also will not build 3D the way a human artist does. If you want hand-crafted models with full control, that is a different craft — the Blender beginner guide covers building scenes by hand in a real 3D editor. This model is for generating code-first scenes quickly; Blender is for designing them deliberately. And if you publish a generated scene, follow the skeptical checklist from the independent explainx.ai review: verify the provenance of the checkpoint you ship, sandbox untrusted generated code before embedding it in a site, and keep a fallback plan, because a preview model's outputs change between checkpoints.

Which Model Am I Actually Downloading? (Avoid the Wrong One)

Three names collide around this release, and downloading the wrong one wastes a 2.71 GB download. Qwen3-Coder is Alibaba's official coding model — a different product with its own install guides; if a search result says "Qwen3-Coder" without "3D," it is not this model. ccore/Qwen3.5-4B-GLB-3D-Coder is the same author's sibling model that generates GLB game assets instead of Three.js scenes — same family, different output. The model this guide covers is ccore/Qwen3.5-4B-3D-Coder: the one whose card says "Early preview · General coding with a focus on 3D" and whose license is Apache-2.0. Check the repo name before you pull.

Sources

Every fact in this guide was verified against a source fetched on November 3, 2026:

• Model card: ccore/Qwen3.5-4B-3D-Coder — base model, 4,146-example dataset, GGUF sizes, generation settings, pinned system prompt, LM Studio MTP warning, duplicate-variable known issue, checkpoint history, FP16 and vision updates
• Hugging Face model API — license (Apache-2.0), download and like counts, last-modified date, linked Spaces
• Model discussion #1 ("Feedback loop") — CCORE's ~60% one-shot success figure and the base-model comparison
• CCORE's Hugging Face profile — author identity and the "Small Models for 3D Creation" collection
• webdiga.com: "Qwen3.5-4B-3D-Coder: Run a 2.7 GB 3D Scene Model Locally" — independent hands-on run: Modelfile, llama serve command, measured CPU-only run, the two real debug fixes, CDN-internet note
• ollama.com/maternion/qwen3.5-3d-coder — community packaging, tags, and install commands
• explainx.ai: "Qwen3.5-4B-3D-Coder: 4B Model Builds Three.js Scenes" — independent fact-check and builder cautions
• Spheron GPU recommender — VRAM estimates and cloud GPU cost figure

Frequently Asked Questions

How big is Qwen3.5-4B-3D-Coder?

The recommended Q4_K_M quantization is 2,708,804,096 bytes (2.71 GB); the near-lossless Q8_0 is 4,482,402,816 bytes (4.48 GB). Those are the exact file sizes on the Hugging Face repo, so budget a few extra gigabytes for the runtime's working files.

Can I run Qwen3.5-4B-3D-Coder on a laptop with no GPU?

Yes. An independent tester ran the full generation on 8 CPU cores with no GPU, using 4.1 GB of RAM, at 9.6 tokens per second — 10,609 tokens in 18 minutes 25 seconds. It works, but generation is slow: start a scene and come back to it, or use a machine with a GPU if you want faster turnaround.

Is Qwen3.5-4B-3D-Coder free?

The weights are Apache-2.0 and free to download, run, modify, and use commercially, and Ollama, LM Studio, and llama.cpp are all free software. No price is published anywhere for the model itself. The only money involved is if you rent cloud GPU time — one GPU calculator lists the cheapest fitting card at about $0.48 per hour — which a local run avoids entirely.

Do the generated scenes work offline?

The model runs fully offline, but the HTML it generates loads Three.js from a CDN link, so the browser needs internet access to render the scene. If you need a truly offline file, download the pinned Three.js 0.160.0 module once and point the page's import at your local copy instead of the CDN URL.

Will the model keep changing?

Almost certainly. It is an early preview with the author still training it — three checkpoints shipped in its first three days, and an October 26 layer fix required re-downloading the GGUFs. Treat any copy older than the October 27, 2026 checkpoint as stale, and re-check the model card before any important project.