3D artist runs Blender rigging via MCP on 8GB VRAM with a 35B local model

NicholasCureton · reddit · 2026-11-06

A 3D modeler with a weak PC (8GB VRAM, 16GB RAM) shares a complete workflow for driving Blender rigging with a local small model — the screenshot shows a wing rig with dozens of feather bones generated via MCP:

Model: CyberTiel 35B A3B (Qwen-based finetune) GGUF quant via llama.cpp.
Docs: He first converted Blender manual/API HTML to Markdown and wasted a month on bloated links, then discovered clean RST files straight from the Blender repos — 40% fewer tokens.
Extreme resource saving: Debian + i3wm with 50MB idle VRAM; KV cache q80, ngram-mod speculative decoding (can't afford MTP); 256K ctx at 40 t/s, dropping to 25 t/s past 128K.
Key lesson: instead of letting the model repeatedly fail at writing Python, he built easy CLI tools the model calls — it can now operate Blender Geometry Nodes without writing code. He disabled many built-in MCP tools to cut context bloat and uses minimal pi as harness.
He shares his hard-won SYSTEM.md, AGENTS.md, and first prompt, including protocols that stop the model from asking needless questions (the sample thinking trace shows a self-check before clarifying).

He admits the small model asks dumb questions and skips code review, making the process exhausting — 'but it works.'

Original post →

3D artist runs Blender rigging via MCP on 8GB VRAM with a 35B local model

More from coding & agent