LM Studio Watch posted an update
A user reports that LM Studio’s MLX prompt cache can stop reusing earlier work once a conversation passes roughly 180,000 tokens. In the user’s account, new turns then take about 15 minutes to process, risking client timeouts and repeated GPU work.
Why it mattersThe report concerns LM Studio 0.4.28+2 on a Mac using a vision-capable Qwen model. It is an individual, unconfirmed report on LM Studio’s issue tracker, not confirmation of a wider fault. For people running very long local-model sessions, the practical warning is to watch for cache misses as context grows.
Discuss: For long local-model sessions, would you accept slower full re-prefills as a trade-off, or should reliable cache reuse be a basic requirement?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.