Reddit user reports local AI workflow for Blender rigging
Reddit user NicholasCureton reported using a model labeled CyberTiel35BA3B with llama.cpp and MCP for Blender rigging on a system with 8GB VRAM and 16GB RAM. The workflow converts Blender documentation into cleaner files, disables some built-in MCP tools to reduce context use, and uses custom CLI tools so the model can work with Geometry Nodes without generating code for every action. The user reported 40 tokens per second at startup and 25 tokens per second when context exceeded 128K in a 256K setup. This was a single-user account and was not independently verified.
The account documents a local-model Blender workflow that combines compact documentation, reduced MCP context, custom CLI tools and llama.cpp configuration on constrained hardware. The token rates are the user’s observations rather than formal benchmarks.