deeboi (accidentally) spent half my token budget convincing claude opus 4.7 (thus the big cost) that spring boot 4 has been released
cheeserwe're (well, I, at least) just use sonnet 2.6 everywhere (home and work) and it's fine.
ParaJust add boot 4 sources into context.
Para"Why did my one prompt cost 2000 dollars?"
ParaI've been happy with Sonnet as well. Step by step only, light agentic work.
deeboyeah my opencode somehow defaulted to 4.7 and i didn't notice, sonnet usually works fine
dreamrealopus took convincing? That's... weird
ParaWhat's the cut-off point of Opus?
deebotried using qwen3.5 locally, but selfhosting this stuff is still randomly just plain broken
ParaWe had a weird withdrawal moment for a few members today because our Enterprise OpenAI subscription was closed.
deebothough when it worked, it was fine, even a 9b model
ParaBasically one stated that his life is now ruined and another that the company can't do this (ummm). We have Copilot available so it's literally just flipping a switch to get the things back but eh, creatures of habit and all that.
ParaOf local models I've been the happiest with that one. Not that I've used it much.
ParaI think Gemma4/Gwen3.5 is overhyped or requires some specific configs I'm too dum to apply.
deeboi bought an intel b50 to test this stuff with, but intel recently changed where they put their support efforts (backend wise), these chinesium models only work with bleeding edge llama.cpp etc
ParaLM Studio indeed uses llama.cpp at its core through their llmester daemon.
ParaSo you can run it with fancy UI or as socket server providing a thingycompatible API.
deeboyeah 3.5 requires some weird hoops to jump through, haven't tested 3.6, and i don't think there's smaller versions of it available
Para(my mind's fried today, have had too many meetings so no proper terminology capacity anymore)
deeboyeah i have it running on a k3s server via the openai api, works ok but the tuning and hassle is barely worth it
deebomaybe when one of the ai companies fails and you can find nvidia h200's in trash bins
ParaMicrosoft apparently has overbought capacity and can't get it to work on Azure.
ParaPlus now that OpenAI is basically leaving them I bet that's going to leave a lot of capacity behind.
ParaWould be kinda amazing if MS was the first company to be displaced by AI.
dreamrealyou mean MS by copilot copilot please god use copilot microsoft windows azure can we fit some more product names copilot?