@vnprc_@valkenburgh@atlbitlab@0xSero@MiaAI_lab To add to this: every business should have an easy way to own their own intelligence and this is what we are pushing towards with Rebase Labs.
I wanted to see if a local model could do real 3D work, so I ran a small experiment.
I asked four models to build a photorealistic 16-inch MacBook Pro in Blender. Same prompt, one shot, Blender MCP only.
Qwen 3.8 27B ran locally on my 128GB M5 Max MacBook Pro through Pi and Ollama. Haiku, Sonnet and Opus 5.5 ran in Claude Code.
Total times: Haiku 5.5: 3m 17s, Sonnet 5.5: 4m 51s, Opus 5.5: 12m 32s, Qwen 3.8 27B: ~1h 16m.
Opus obviously produced the best-looking render, especially on form and lighting. Sonnet came in second.
Haiku and Qwen were a toss-up, since each fell short in different ways. Qwen couldn't get the lid open to the 110° the prompt asked for, didn't orient the screen correctly, and got the keyboard proportions wrong. Haiku's overall body shape was off, and the lighting in its final render was poor.
I plan to run more tests in the future, but if you've tried this, let me know how it went, and share any models and/or setups that have worked well for you.