GLM-5.3's reasoning_effort silently defaults to max, including when you typo it

DEV Community (model cards and Unsloth GGUF pack dates cited as primaries)·medium signalXBlueskyLinkedInCopy link

A practitioner writeup measured what it actually takes to run Z.ai's GLM-5.3 (753.9B, weights out August 28) and GLM-5.3-Flash (320.8B with 18B active, out August 27) locally. Smallest usable quants are 216.7 GB for the flagship and 93.1 GB for Flash, and counterintuitively the 753.9B flagship runs in stock llama.cpp while Flash does not yet. The trap: `reasoning_effort` takes low, high and max, and the model card says it defaults to max if unset or set to any other value, so a typo buys a long think on every message. Through a hosted OpenAI-compatible provider the behavior differed, with `medium` honored as a real middle rung and an omitted field landing at 30 output tokens rather than max's 48.