GLM-5.2 local-inference quantizations documented

Category: AI
Status: discussion
Sources: unsloth docs, weights, discussion
Summary: Unsloth published local-deployment guidance for the open-weight GLM-5.2 (753B-parameter MoE, MIT license). It lists dynamic GGUF quants: a 2-bit UD-IQ2_M build of about 239 GB that the docs say fits a 256 GB unified-memory Mac or a single 24 GB GPU with 256 GB system RAM and MoE offloading, and 4/5-bit builds described as near-lossless. Reported accuracy figures (76.2% at 1-bit, 82% at 2-bit) are the project's own, not independently verified.
Why it matters: It sets concrete memory floors for running a top-scoring open-weight coding model locally, which matters for teams weighing GLM-5.2 against metered proprietary APIs.
Follow-up: Watch for independent local-throughput and quality measurements of the 2-bit and 4-bit GLM-5.2 quants.