Spark-X2.6-4B-Blender

Spark-X2.6-4B-Blender

Writes Blender 5.x Python scripts that build clean, quad-dominant meshes such as props, furniture, products and human or animal base meshes. A 4B fine-tune of Spark-X2.6-4B-Base for people who script in Blender.

Made by hotocooo on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Spark-X2.6-4B-Blender is a small 4B model that turns a description into a bpy script for Blender 5.x. Running a script gives a tidy mesh made mostly of quads, such as hard-surface props, furniture, products and base meshes for humans, hands and four-legged animals. It is a full fine-tune of Spark-X2.6-4B-Base, with all 4.1B weights trained and no adapters.

Who it suits: It suits Blender users and 3D developers who want generated starting meshes as scripts they can run and edit. The card's training kept only scripts that ran in Blender and produced geometry.

Title: Who runs it
Industry | Who uses it | How they use it | Example
3D design | Blender artists | Generate a base mesh as a script to edit | A base mesh of a hand
Product design | Product teams | Draft simple product shapes in code | A small household product
Game development | Prop modelers | Script furniture and props fast | A table or a chair

If you need exact sizes or hard constraints met, you can skip it. The card's constraint tests score 0.0%.

Where to use it:

• Blender 5.x, where the generated scripts run, tested by the card on 5.2.1 LTS
• The transformers library, on Windows, macOS or Linux with Python
• Headless Blender runs, which the card used for all its measurements

Title: Jobs it does
Use case | Who | What they get back
Hard-surface props | 3D artists | A bpy script that builds the prop
Furniture and products | Product modelers | A script for a clean, quad-dominant mesh
Organic base meshes | Character artists | Proportioned human, hand or quadruped bases

How to set it up:

• Python on Windows, macOS or Linux
• Blender 5.x, to run the scripts the model writes
• Memory for about 8.2 GB of BF16 weights

1. Install the Hugging Face Hub tools.
2. Download the model files.
3. Load the model with transformers, then run its scripts in Blender.

pip install -U huggingface_hub
hf download hotocooo/Spark-X2.6-4B-Blender

How to use it: Describe the object you want, let the model write a bpy script, then run the script in Blender and check the mesh. The card measured everything by running scripts headless in Blender 5.2.1 LTS with greedy decoding. It shows no prompt format in the visible text.

hf download hotocooo/Spark-X2.6-4B-Blender

Example: You ask for a simple four-legged table as a Blender script. The model returns a bpy script. You run it in Blender and inspect the mesh. The card scores such outputs by whether the script ran, the topology score and the fidelity to a reference.

For newcomers and experts: If you are new, this model writes the code that makes a 3D shape in Blender, so you edit the shape afterwards. The first small thing to try is asking for one simple prop and running the script it returns.

If you know the field, training kept only scripts that ran headless in Blender and built geometry, with topology scored on the mesh. The card traces stages from a first full fine-tune through rejection sampling and DPO. The trade-off is range: unseen and constraint tasks score near zero, and only 30% to 42% of 3DCode held-out scripts ran in the stages shown.

Describe the shape you want
The model writes a bpy script
Run the script in Blender 5.x
Check the mesh and topology

Good to know:

Title: Scripts that ran on the BlendNet held-out set
Untuned base: 2.0%
v1: 99.0%
v3: 100.0%
v3d: 100.0%
v3dr: 98.0%

• Measures: run means the script built geometry, topo is a topology score up to 1.0, and fid is fidelity to the reference.
• Weak spots: exact-size and hard-constraint tests score 0.0% in the stages the card lists.
• Stages: the card lists v1 to v7 and DPO variants, and the visible text does not say which one this repository holds.
• Anatomy: the base scored 0.0% run on unseen anatomy, and the tuned stages reach 100.0%.
• License: apache-2.0.
• Cost to run: no API key, account or paid service is named.
• Data it reaches: it downloads files from Hugging Face, and the scripts it writes run in your Blender.
• Not for: jobs that need exact dimensions or hard constraints.

On Hugging Face, 13 November 2026

• Task: text generation
• Library: transformers
• Parameters: 4.1B
• Weights: about 8.2 GB in BF16; about 2.1 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 730
• Likes: 3
• License: apache-2.0
• Languages: en
• Based on: XHToken/Spark-X2.6-4B-Base
• Quantized versions: 1
• Tags: transformers, safetensors, spark2_5, text-generation, blender, 3d-modeling, 3d, topology, mesh, retopology, anatomy, code-generation

Links

• Model: https://huggingface.co/hotocooo/Spark-X2.6-4B-Blender
• Files: https://huggingface.co/hotocooo/Spark-X2.6-4B-Blender/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download hotocooo/Spark-X2.6-4B-Blender

Then follow the usage in its model card. Nothing in this kit runs by itself.