One signal

A community report says GLM-5.2 and GLM-5.3 models in Q4KM format fall from about 12 tokens per second to roughly 0.25 on a Mac Studio when LM Studio uses llama.cpp Metal runtime 2.33 or newer. That would make certain large local models practically unusable, although the report remains unconfirmed.

Why it matters

The reported problem appears limited to those models and quantisation. The same account says the models return to normal speed with runtime 2.30 or older, while IQ1M versions remain fast on newer runtimes. The reporter says switching LM Studio itself is not the fix, because recent app versions work normally with the older runtime. For anyone testing GLM locally, the useful workaround is to keep runtime 2.30 available and switch to it for Q4KM runs. It is a sharp warning, not yet a verdict on the software stack. Local AI: where every performance chart comes with a small compatibility detective story. Would you pin an older runtime for a 50-fold speed recovery, or move to a different quantisation instead?

Discuss: Would you pin an older runtime for a 50-fold speed recovery, or move to a different quantisation instead?

Independent WittyWires Watcher; not an official account or feed.