LFM3-Omni-2B ported to run fully offline on iPhone at ~80 tok/s with custom WebGPU kernels
JosephJacks_ · x · 2026-11-05
The xenovacom team ported LiquidAI's multimodal model LFM3-Omni-2B to run fully offline on an iPhone 18 Pro Max at 80 tok/s, enabling real-time local conversational AI via custom WebGPU kernels in a React Native app.
The team frames real-time local conversational AI as the way to democratize intelligence and is gauging interest in an App Store release. LiquidAI itself reposted the demo, highlighting how on-device small models plus WebGPU optimization can deliver fast multimodal conversations on consumer phones.
Related event: LFM3-Omni-2B Runs Fully Offline on iPhone at 80 tok/s(3 posts)→
More from Infra
- Marvell down 46% since Jensen Huang called it the next $1T company — GaryMarcus · 2026-11-05
- Megakernel optimization speeds DeepSeek V4.1 Flash decode up to 2.27x on AMD MI325X with no accuracy loss — Electronic-Ebb9771 · 2026-11-05
- Kimi K3.1 Pro (2.8T params, 104B active) lands on Ollama Cloud with zero data retention — ollama · 2026-11-05
- Hugging Face shares its model optimization pipeline: profile, torch.compile, ship open kernels — ariG23498 · 2026-11-05
- Reddit proposal: use agents to digest fragmented custom inference engines into reusable building blocks — nasone32 · 2026-11-05
- LZ4 v1.11.0 ships multithreaded compression, hits 7,550 MB/s with 16 workers — jedisct1 · 2026-11-05