AioLM v0.1.17

Windows installers are unsigned. Verify the SHA-256 values in checksums.txt from this release.

What's changed

• Preview VRAM, RAM offload and SSD offload before loading a model. Estimates update with settings and explain capacity limits, unsupported layouts and assumptions; stored model file sizes are shown separately.
• Make inherited defaults and customized profile settings easier to distinguish. Profile details group GPU placement and runtime arguments, and reset controls sit beside each setting's title.
• Allow desktop, CLI and benchmark model startup up to ten minutes for loading and warmup, while reporting process exits promptly and preserving cancellation.
• Simplify the chat composer and show the selected model in conversation session details.
• Add first-run language, theme and model-folder setup. Keep app data together in the user data folder and save conversations as readable files.
• Improve profile save comparisons, single-instance startup, tray restoration, uninstall data cleanup and corrected multi-GPU ROCm runtime packaging.

Resource estimates are planning guidance. Firmware-reserved or shared GPU memory and custom multi-GPU allocation ratios can reduce usable capacity compared with the displayed totals.

Validation

• Frontend contracts, direct regression tests and 1,170 Vitest tests passed, alongside type checking, lint and production build.
• Rust formatting, Clippy and automated tests passed. Live model and runtime-download checks remain opt-in and were not run.
• Windows installers are built and verified by the release workflow.

Full changelog: v0.1.16...v0.1.17