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)

LFM3 is now in Private Mind, our open source app for running models locally on iOS and Android. In the video, LFM3-VL-2B reads a Japanese ramen menu: asked what to order without spice, it skips the bowl marked 激辛 (extra spicy) and picks the miso ramen; asked about paying by card, it finds the "cash only" line under the prices. Numbers on an iPhone 16 Pro: LFM3-1.7B (text), MLX 4-bit on the GPU: ~40 tok/s, first token under 0.5 s LFM3-VL-2B with a photo: ~9 s to the first token, ~20 tok/s decode. A text follow-up in the same chat starts in ~5 s Exporting LFM3 to ExecuTorch LFM3 isn't LFM2 with more parameters, and ExecuTorch's existing LFM2 export doesn't cover it. The new pieces: 24 layers mixing gated short convolutions (width 7), full attention and sliding-window attention (window 256) a sigmoid gate on every attention head's output, and RMSNorm on Q and K different RoPE per layer type: full attention rotates only half the head dims with theta 1e6, sliding layers use theta 1e4 sandwich norms around every block So I rewrote the decoder in plain PyTorch for export and checked it against Liquid's reference implementation: max logit difference 0.00005, same top 10 tokens. The part I liked most is the sliding-window cache. ExecuTorch's built-in ring-buffer cache is sized at 2x the window, which caps a single forward call at 512 tokens, so a long prompt has to be fed in small pieces. I gave the sliding layers a buffer of window + 2048 and let the attention mask enforce the 256-token window, so one call can take 2048 tokens and the context still goes to 8192. Two builds came out of it: MLX 4-bit for the Apple GPU (1.07 GB) and XNNPACK 8da4w for CPU on iOS and Android (1.17 GB). The LFM3 exports are on Hugging Face: https://huggingface.co/software-mansion/LFM3-ExecuTorch