v3.2.7 is out, and it is entirely fixes - most of them from reports posted right here on this board. Hit Update in Pinokio.
Install now fails fast on macOS 13 instead of failing weirdly. mlx 0.31.1 ships no macOS-13 build, so on Ventura the engine quietly failed to install and you only found out at the very end with a ModuleNotFoundError - under an auto-generated title blaming Python 3.11, which pointed at completely the wrong thing. Phosphene now checks up front and tells you to update to macOS 14+. (Thanks @hottboytank.)
The "install Qwen-Image-Edit" button that was never a button. In the Image tab that red pill looked clickable, but it was a status indicator - clicking it gave you a question-mark cursor and did nothing. Worse, the hint next to it pointed at a sidebar item that had since been renamed. Both fixed: it now reads "Qwen-Image-Edit not installed" and tells you the engine pack ships with Update. (Thanks @poppy0396 - sorry that one sat far too long.)
Character LoRA training that silently produced nothing. The trainer forced the dev transformer, but only checked the file existed - not that it was full precision. So a 4-bit quantized dev (~11 GB) sitting in the Q4 folder got loaded, and training against quantized weights produced a near-zero LoRA that applied to nothing. It is now size-aware, prefers the full ~19 GB dev, and warns loudly when a quantized one is shadowing it. (Caught by saved-j on GitHub, who did a lot of unpaid debugging to get there.)
Metal GPU timeouts now say what they are. A crash previously reported as "C-level assertion failed" is usually the macOS GPU watchdog killing an over-long Metal command buffer during prompt encoding. It now names itself instead of sending you hunting the wrong thing.
Also: the stats-dashboard 403 no longer spams the log looking like a real failure.
Not in this release: the Remix wave - Ingredients, Control, Colorize, and running a trained character LoRA on top of ingredients. That is a much bigger change that pulls new model weights, and it gets a full fresh-install validation pass before it ships. I would rather it land properly than land fast - the last fresh-install manifest bug cost some of you weeks.