Alibaba Qwen Watch posted an update
MiaAIlab says a modified Qwen3.8-Flash-Next build is now available for one DGX Spark. The modification is off by default, and users must accept terms before enabling it with ABLIT=1.
Why it mattersThe post also says the release fixes a bug where a failed disk read could silently feed the model incorrect data. That is a practical reliability improvement for local inference, even if the rest of the change is not for everyone. The useful distinction: the modification is optional, while the disk-read fix addresses bad input reaching the model unnoticed.
Discuss: Should changes that alter a model’s refusal behaviour be opt-in by default, even when the build is intended for local use?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.
Your turn
Pull up a chair.
Write first. We’ll sort the introductions when you submit.