Reviews ยท 4 written up
Hardware runs the same workload set for ninety days before it gets a score. Software gets tested in a tenant with real users in it. Nothing here is sponsored and nothing here is a referral link.
Fast, complete and officially maintained. The constraint is not speed, it is that one snapshot of a content-heavy page can eat 31,800 tokens of context.
Boots in under five seconds and idles at 140 MiB, a fraction of anything else here. Transcoding on a laptop CPU tops out around two 1080p streams.
Read all 120 fields off 24 synthetic scans without a miss, and idles at less than half what Immich needs. The easiest self-hosted stack here so far.
Installs in half a minute and never wobbled. It is also not a lightweight photo app: 2.42 GiB idle and it will take every core you own.
Method
Hardware gets three nodes, one sustained backup window and a deliberately hostile firmware update. Devices that fail the update test do not get a score at all, they get a paragraph explaining what happened. Software gets a real rollout: rings, a rollback plan, and a note on what the admin experience is like at 2am.
No. Nothing on this site is sponsored and there are no referral or affiliate links. Review units are bought or loaned, and loaned hardware goes back within thirty days of publication with a note on the review page saying so.
At least ninety days in a working 4U rack, running the same workload set every time: three compute nodes, a sustained backup window, and one deliberately hostile firmware update. Power is logged at the wall and noise measured at one metre.
Software gets a real rollout in a tenant with real users, in rings, with a rollback plan. The score reflects what the admin experience is like at 2am, not what the feature list promises.
It does not get a score. It gets a paragraph explaining exactly what happened and at which stage, which is usually more useful than a number.