LFM3 on iPhone for the PWA
Lead: Safari on this iPhone only loads the smallest Liquid model. LFM2.5-350M runs at 40+ tokens a second; anything 760 MB or bigger crashes the tab. A native server on the phone avoids that limit.
Safari test results
| Model | Size (4-bit) | Result || LFM3-1.7B | 1.38 GB | no tab crash at load || LFM2.5-1.2B | 760 MB | no tab crash at load, cache off too || LFM2.5-350M | 255 MB | ok 9 s load, 39–48 tok/s, weak answers |An API served from the phone
The catch
iOS pauses normal apps in the background. The Phone Link helper already stays alive in the background, so it is the best host. Device builds need the newer Xcode that comes with the macOS 27.2 update.