The local Qwen 3.8 27B might not be anywhere near as smart as Claude's models but I never wake up one morning and wonder if I am part of some test model to see if I will accept something somewhat less capable or not trained as well in the tools I am using day to day. Its also not suddenly going to start refusing to test the security of my system or scanning the code for vulnerabilities or continuing to back engineer that old binary the company lost its code for years ago.

It is something I also appreciate with some of the other labs like deepseek who publish the checkpoint they are running on and still sell you the old one for a period of time as well.

Its obnoxiously anticonsumer to be constantly chopping and changing what a model represents on customers who are paying them an awful lot more to be on the frontier models.