Switching the primary LoRA in Studio no longer reloads the model. Choose another primary and press Render: the old LoRA comes off and the new one goes on the model already loaded, in a second or so, where it used to reload the whole model (minutes on a large one such as MOSS-Image). The render is the same as a fresh load of that LoRA. A LoRA that can't load (one for another model, say) leaves the old one on and says why. MiniMax H3 video and Frankenstein mode still load again. (#394, thanks @tsk5cb-cmyk)
AMD: MIOpen finds its packaged kernel database. Both AMD launchers pointed MIOpen at a folder with no database in it, so it couldn't use its packaged kernel tuning. The Linux launcher now points at the database (#384, thanks @xGnome), and so does Launch Fizgig AMD.bat on Windows (#389, thanks @dkspwndj for the report).
A GGUF file no longer slows Fizgig down. A GGUF quant in the secondary captioning model slot was read in full at every page load, about 15 seconds for a 20 GB file; the captioner now takes only .safetensors files (#388, thanks @spaghettified26). Every other place that reads a model file's header refuses a file like that from its first eight bytes too, where it used to read gigabytes before failing.
To update: run the updater as usual.