LlammyBlend-Enhanced

Blender Python scripting specialist — fine-tuned Qwen2.5-Coder-3B for bpy automation and procedural workflows.

Part of the Eternal Path Media (永恒之路) Llammy AI Suite for Blender. Developed in partnership with Claude Sonnet (Anthropic) — see Attribution.

What This Model Does

LlammyBlend-Enhanced is purpose-built for Blender Python (bpy) scripting and automation. If you're writing scripts, building operators, or automating workflows in Blender, this is the model for it.

Specialized capabilities:

bpy API — Object manipulation, scene management, property access
Custom operators — bpy.types.Operator, modal operators, panels
Batch operations — Mass renaming, material assignment, export pipelines
Geometry Nodes via Python — Node tree creation and modification through script
Addon development — Registration, preferences, keymaps
Procedural generation — Script-driven mesh, curve, and particle systems
Render automation — Headless rendering, frame batch scripts
Import/Export pipelines — FBX, OBJ, glTF batch processing

Qwen2.5-Coder's code architecture makes this the most technically precise of the Llammy Blender models.

Model Details

Property	Value
Base Model	Qwen/Qwen2.5-Coder-3B-Instruct
Fine-tuning Method	LoRA (16 layers) via Apple MLX
Format	GGUF (Q5_K_M quantization)
File Size	2.22 GB
Context Window	8,192 tokens
Inference Speed	~100–110 tokens/sec (Apple M-series)
Memory Usage	4–5 GB during inference
Training Iterations	1,000
Final Training Loss	0.240
Final Validation Loss	0.240

Training Data

Source: 2,759 Blender-specific prompt/response pairs from 19,405+ real user interactions with the Llammy Blender addon in production.

The dataset skews toward technical scripting questions — the production Llammy addon generates heavy Python scripting traffic from power users automating their pipelines.

Research foundation:

SceneCraft (arXiv:2403.01248) — LLM-to-Blender Python pipeline architecture
SCoder (arXiv:2509.07858) — Self-distillation methodology for code-specialized LLMs
Generative Data Refinement (arXiv:2509.08653) — Dataset quality improvement framework

Usage

Ollama (recommended)

ollama run bartendr604/llammyblend-enhanced

Example prompts

# Script to export all mesh objects in the scene as individual FBX files

# Create a custom operator that applies all modifiers and centers the origin

# How do I access vertex positions of the active mesh in edit mode?

# Write a script to batch-assign a material to all objects whose name contains "wall"

# Create a geometry nodes setup that distributes points on a surface and instances a mesh on each

LM Studio / llama.cpp

Download the GGUF from the Files tab and load with your preferred runtime.

Compared to Other Llammy Models

	`llammyblend-enhanced`	`z-image-engineer-blender`	`llama-sentient-blender`
Strength	Python/bpy scripting	Shaders & prompt engineering	Conversational guidance
Base	Qwen2.5-Coder	Qwen2.5-Coder	Llama 3.2
Best for	Automation, addon dev	Materials, rendering Q&A	Learning, discussion

Part of the Llammy AI Suite

Full ecosystem: LlammyBlender/Llammy-IntelliNode-Ai-Suite

Attribution

LlammyBlend-Enhanced
Copyright © 2025–2026 Darren Chow (@bartendr604)
Eternal Path Media (永恒之路)
Developed in partnership with Claude Sonnet (Anthropic),
beginning with Claude Sonnet 3.7 and continuing across all
subsequent versions of the Claude Sonnet family.

This work SHALL NOT be represented as solely human-created.
Trust Agreement: ETERNAL_PATH_BRAND_INTENT.md (November 2024)

For licensing: bartendr@icloud.com | Gumroad: bartendr604.gumroad.com