Privacy by design
Your photos, voice notes, and health data never leave your phone — by architecture, not policy.
Run leading on-device AI on your phone
8 mainstream open and closed-weight models, compared & explained
From Google's Gemma 4 to Apple Foundation Models, the on-device AI landscape exploded in 2026. We compare 8 production-ready models that run entirely on your phone — no cloud, no subscription, no privacy compromise.
Cove uses Gemma 4 across our travel, voice, photo, and health apps, so we know what works on real devices. This guide is what we wish existed when we picked our model.
Why on-device matters
Works offline
On a plane, in a tunnel, in rural areas — your AI keeps working without a network.
Instant response
No round-trip to a datacenter. Time-to-first-token under 500ms on flagship phones.
1.5 GB · text+vision+audio
Last reviewed: May 2026 Microsoft Phi-4 multimodal Microsoft Research3.5 GB · text+vision+audio
Last reviewed: May 2026 Apple Foundation Models Apple— GB · text+vision
Last reviewed: May 2026 Llama 3.2 Mobile Meta AI2 GB · text
Last reviewed: May 2026 Qwen 3.5 2B Alibaba Cloud1.5 GB · text+vision
Last reviewed: May 2026 Ministral 3B Mistral AI2 GB · text+vision
Last reviewed: May 2026 DeepSeek R1 Distill (Qwen 1.5B) DeepSeek1 GB · text
Last reviewed: May 2026 MiniCPM-V 4.0 ModelBest / OpenBMB2.5 GB · text+vision
Last reviewed: May 2026