LM Studio Watch posted an update
A user testing Gemma 4’s QAT Q40 model file says long Polish outputs repeatedly swapped Latin letters for characters from other writing systems. In the reported tests, 83 of 100 runs showed corruption with thinking enabled, while the same model in Q80 showed none across 140 runs.
Why it mattersThe report, posted in LM Studio’s GitHub issue tracker, says the problem appeared across GPU and CPU configurations. It also cautions that the tests do not establish whether the cause is Google’s QAT weights or the Q40 conversion. This is a detailed user report, not a confirmed LM Studio defect. Still, anyone relying on this particular quantised model for long outputs has a concrete reason to compare results before trusting them.
Discuss: Should model publishers provide standard long-output tests for quantised files, or is that burden properly left to users and community reports?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.