This month I reduced my spend on cloud AI and started using local models as primary - it's maybe 2 or 4 times slower for some things but it still is getting the job done just fine. I talked to a friend of mine who's been doing the same running smaller Qwen models locally for coding.