Deepseek 4.1 Flash 1017 as what i did the latest demo with, GLM 5.3 flash also works well for driving a 3d accelerated desktop. I am always experimenting with the latest releases to see how they do.
The stack of 5090s are for smaller workloads, the 4x6k’s run the big reasoning model, i also have tts models, embedding models, instruct and special vision models… the rig is a busy place.
Proxmox is running on a different machine for the 3d desktop VMs using older ampere and ada generation rtx 4k’s doing simple pcie passthrough.
Developing RCON tooling to allow multi-turn multi-input tool calling was a mind bender. How do you let the LLM hold control, and look at the screen, then press tab until it selects the right option, then release control? Clicking and dragging to move 3d models around is where my tests with video embedding come in. Returning multiple multimedia (video/image) from a single tool call is something i had to patch into turnstone.