GLM 5.2 on a Mac Studio M3 Ultra: 395GB, 12 tok/s, 74K Context

Key Takeaways

• Hardware: M3 Ultra Mac Studio with 512GB unified memory; GLM 5.2 is a 395GB download that lives almost entirely in RAM.
• Performance: ~12 tokens/sec generation and roughly 74,000 tokens of context before the machine hit its limit.
• Runtime: the model was too new to run in LM Studio natively, so it needed a custom runtime to load at all.
• Claude Code timed out: at ~12 tok/s, generation was slow enough that requests kept timing out inside Claude Code — so he demoed in Pi Agent inside Cursor instead.
• The harness matters more than the model: a basic prompt to Pi Agent gave a basic app, while a strong harness like Claude Code turns a simple prompt into a far more complete result. Pair a local model with a good harness when you can.