LFM3-1.7B and LFM3-VL-2B running fully offline on iPhone and Android, plus what it took to export LFM3 to ExecuTorch (open source app)

Details

LFM3-1.7B and LFM3-VL-2B are now available in Private Mind, an open source app for running models locally on iOS and Android.
On an iPhone 16 Pro, LFM3-1.7B (text) achieves ~40 tok/s with MLX 4-bit on the GPU, with first token latency under 0.5 s.
LFM3-VL-2B processes a photo in ~9 s to the first token and decodes at ~20 tok/s; text follow-ups start in ~5 s.
The export required rewriting the decoder in plain PyTorch because LFM3 architecture differs from LFM2, featuring 24 layers with mixed gated short convolutions, full attention, and sliding-window attention.
Custom sliding-window cache implementation allows 2048-token context in a single forward call, expanding total context to 8192 tokens.
Two builds are provided: MLX 4-bit for Apple GPU (1.07 GB) and XNNPACK 8da4w for CPU on iOS and Android (1.17 GB).
Exports are hosted on Hugging Face at software-mansion/LFM3-ExecuTorch.

Key point