GLM 5.2 local setup with Hermes: hardware and tradeoffs

The speaker describes running GLM 5.2 on a single Mac Studio through Hermes Agent. His demonstration is a 3D first-person shooter that he says the agent built, tested and improved. He compares the result with his Opus 4.8 test, but this is a report of his own experiments rather than an independently verified performance comparison.

Hardware is the main constraint. The speaker uses Unsloth's two-bit version, which he describes as roughly 250 GB in size and requiring about that much memory. He warns that a 256 GB Mac Studio leaves little room and may crash; he recommends a 512 GB configuration for this workload. For smaller machines, he suggests Gemma 4 or Nemotron, and names Qwen 3.6 27B as an option for midrange hardware.

The setup section follows an AI agent workflow: he gives Hermes a link to Unsloth's announcement and asks it to install the model on another Mac Studio. He says Hermes handles the download and server setup. This is an account of automated installation, rather than a command-by-command tutorial.

For this local LLM, the speaker favors background code review and bug fixing over interactive coding. He reports slow responses and a smaller context window compared with the cloud models he uses. He presents keeping prompts on his computer and avoiding per-token cloud charges as reasons to run models locally, while acknowledging hardware and electricity costs.