imference-desktop Files

✨ MiniMax H3 video generation

v0.4.0 adds support for MiniMax H3

Keep your models

Switching models used to delete the previous checkpoint. Downloads are now kept on disk under a size quota (100 GB by default) — coming back to a model is an engine restart, not another 7 GB. When room is needed, the least recently used ones go first, never the one you're using and never your own checkpoints.

A Storage screen in Settings shows what's cached and what it costs, deletes a single model, purges the shared base components, or opens the folders.

Before a download starts, the app checks the file actually fits — with a free space cushion — instead of evicting models and then failing anyway.

One click to generate

A model already on disk now shows Generate (or Add to queue), not Download. Loading it into the engine happueued. Downloads of several GB remain a deliberate, separate click.

Queued jobs are tied to the model they were written for, so adding to the queue while another model is running can't hand y

Reference images

Optional reference images, up to two slots d to the model's working resolution before they're sent. They now live in their own card, outside the prompt box.

Models that can't use one don't offer one —all, and dropping an image on such a model says so instead of silently ignoring it. Cloud reference images are off for now,.

Create panel

• Format is a proper grid: each option d its output resolution in pixels, sorted from widest to tallest.
• Parameters match the model. A settinga value instead of a slider that pretends to move, and one it doesn't have at all simply isn't there.
• Clip length is selectable on video models that support it.
• The Generate button floats at the botrolling back up to launch a run.

Gallery

• Sound control on videos: one tile plays audio at a time, so unmuting one mutes the others. The redundant "this is
• Videos pause when off screen and release their memory once far away, which ends the dozens of clips that kept decodi
• Long errors in the queue are readable in full, and selectable.

Cloud

• The price shown next to Generate now reflects your format and clip length, and matches what is charged.
• A busy queue no longer looks like a failure: a run that outlives its polling window stays pending and is picked back u
• Out of USDC says so, with the amounts, instead of a generic server error.

Under the hood

• Local engine upgraded to imference-engine v0.3.4 — model swaps use far less host RAM. The venv upgrades itself otself** if a previous upgrade was interrupted (which left it unable to generate).
• Built on Wails v3 beta, with Windows
• Two dev-only dependency advisories patched (vite, postcss).

Full Changelog: https://github.com/Publikey/imference-desktop/compare/v0.3.5...v0.4.0