Gemma 4 on iPhone
deepmind gemma google multimodal
| Source: HN | Original article
Google DeepMind’s Gemma 4 has landed on iPhone, marking the first time the frontier‑level open‑source model can run entirely on iOS hardware. The rollout arrives through Apple’s Core ML framework and third‑party wrappers such as Novita AI, which now expose all four Gemma 4 sizes – the on‑device‑friendly E2B and E4B, plus the larger 26‑billion‑parameter and 31‑billion‑parameter variants – to iPhone 15 series and later devices.
Gemma 4 expands on its predecessor, Gemma‑3 n, by adding multimodal capabilities: it accepts image, text and audio inputs and can generate text, summarize videos, produce study notes, draw simple graphs and even issue commands to other apps. The model’s open licensing means developers can embed it directly into apps without routing data through cloud services, a shift that promises lower latency, offline operation and stronger privacy guarantees.
The move matters because it challenges Apple’s own on‑device language models and the broader industry’s reliance on proprietary APIs. As we reported on 5 April, Gemma 4 delivered “frontier‑level performance” on a 48 GB GPU, outperforming many closed‑source rivals in benchmark tests. Bringing the model to iPhone demonstrates that the same performance tier can be approached on consumer‑grade silicon, potentially reshaping the AI app ecosystem in the Nordics and beyond.
What to watch next: early benchmark data from independent testers will reveal how the E2B and E4B variants handle real‑world prompts on the A17 Bionic chip. Apple’s upcoming iOS 18 beta may include deeper Core ML optimisations, and developers are likely to experiment with on‑device assistants, translation tools and creative utilities powered by Gemma 4. Keep an eye on whether Google expands the model‑API pricing or opens additional fine‑tuning tools, and how competitors such as Meta’s Llama 3 respond to an open, multimodal model now native to iPhone.
Sources
Back to AIPULSEN