Two things you’re probably running into with Qwen/Gemma: their raw MCP-tool-call reliability, and needing a much more explicit system prompt than chat use requires.
On model choice, Meta’s Muse Glimmer 30B (released August 2026, Apache 2.0) is specifically strong here, it’s the best MCP tool-use score of any model runnable on one consumer GPU (75.5 on the MCP-Atlas benchmark, vs Qwen3.6-27B’s 62.5), so if tool-calling reliability is your actual bottleneck rather than raw reasoning, that’s worth trying over Qwen/Gemma directly. If 30B is too heavy for your hardware, GLM’s smaller variants also advertise native MCP support with structured output, which tends to matter more for CAD tool calls than general capability.
For settings/system prompt in LM Studio: the two things that fix most “model doesn’t use tools well” problems are (1) explicitly listing the exact tool names and required parameter shapes in the system prompt rather than trusting the model to infer them from the MCP schema alone, and (2) low temperature (0.1-0.3), tool calling wants deterministic, low-creativity output, default creative-writing temperatures make models invent parameters or skip required fields. For Blender specifically, the BlenderMCP addon’s own docs (yuri-schmaltz/blender-mcp) include a working example system prompt structured around “describe the scene state, then call one tool at a time,” worth copying that structure rather than writing from scratch, since a lot of tool-use failures come from a model trying to chain multiple CAD operations in one response instead of waiting for each result.