Breaking

BlenderLLM Released: Open Model Turns Natural Language Into Blender CAD Scripts

July 10, 2026 at 16:57 EDT

Open Source
Foundation Models
Software Dev & Coding

An open-source model called BlenderLLM has been released that automatically generates Blender Python API (bpy) scripts from natural-language instructions to build 3D CAD models. Developed by the research group FreedomIntelligence, it publishes the model weights, training data and an evaluation benchmark, and can run locally.

Open-Source Release · FreedomIntelligence

BlenderLLM: Talk to Build a 3D Model

A locally-runnable, Apache-2.0 model that turns plain-language instructions into Blender Python (bpy) scripts — then executes them to generate and render CAD models. Built on Qwen2.5-Coder-7B, fine-tuned for CAD.

~12K

BlendNet training samples
2K manual + 10K via GPT-4o

700

CADBench eval samples
500 simulated + 200 real-world

3.4%

Syntax error rate
on simulated benchmark

CADBench-Sim Average Score

Judged by GPT-4o — higher is better. BlenderLLM leads the field.

0.748

BlenderLLM

0.687

o1-Preview

0.593

Claude-3.5

0.589

GPT-4-Turbo

0.565

GPT-4o

0.519

BlenderGPT

Real-world set (CADBench-Wild): 0.664 avg · 3.5% syntax error rate.

How It Works

Natural language

describe the object

→

bpy code

model generates script

→

Blender render

execute & build 3D CAD

Strengths

Highest accuracy & lowest syntax errors on benchmark
Runs locally — no data sent externally (privacy)
Open weights, training data & benchmark (Apache-2.0)

Limitations

Basic modeling only — no materials, finishing or internal structure
Text-only input; no images / multimodal, no multi-turn dialogue
Real-use quality reported to fall short; large-scale eval pending

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.