Google Deploys Gemma 4, Its New Open‑Source Model Running Offline on iPhone

Google’s on‑device AI push finally feels tangible. By simply installing the free “Google AI Edge Gallery” from the App Store, iPhone users can pick a Gemma 4 variant and run inference without ever touching a cloud endpoint, according to GizmoWeek. The app bundles a text console, image‑recognition, voice interaction and a “Skills” plug‑in system, turning a phone into a miniature AI lab rather than a one‑off demo. The smallest E2B model, a 2‑billion‑parameter version, is the default because it fits comfortably within the memory and thermal envelope of a typical iPhone, while still delivering the reasoning chops that Google claims stem from its Gemini 1.5 Pro research.

Gemma 4’s technical pedigree is where the magic happens. As Jubin Soni explains in his deep‑dive, the model inherits the “density of intelligence” philosophy of Google’s Gemini line, packing state‑of‑the‑art reasoning, coding and multilingual abilities into a leaner footprint. Its transformer decoder backbone is augmented with Multi‑Query Attention (MQA) and Grouped‑Query Attention (GQA), which slash memory usage and speed up inference. A sliding‑window attention (SWA) layer lets the model handle longer contexts without exploding compute costs, while logit soft‑capping stabilises training and improves output quality. Those tweaks let the 31‑billion‑parameter flagship hold its own against Qwen 3.5’s 27‑billion‑parameter rival, even though Gemma carries roughly four billion more parameters, GizmoWeek notes.

The real surprise, however, is how the mid‑size E4B (4‑billion‑parameter) and the ultra‑light E2B models have been engineered for mobile deployment. Google’s engineers trimmed the architecture to prioritize efficiency over raw capability, resulting in faster, cooler inference that respects the iPhone’s battery limits. In practice, the E2B can generate a paragraph of text in under a second on an iPhone 15 Pro, a speed that rivals many cloud‑based APIs once network latency is factored in. Soni’s analysis points out that these variants still benefit from the same distillation pipeline that powers the larger Gemma 4, meaning they inherit a substantial portion of the larger model’s knowledge despite their modest size.

From a developer’s perspective, the rollout is as simple as a standard App Store download. Once the Edge Gallery is installed, users select their preferred Gemma 4 variant, download the weights (a few hundred megabytes for E2B), and start prompting the model locally. No API keys, no data leaving the device, and no subscription fees. The platform also exposes a Python‑compatible runtime, letting hobbyists and researchers prototype on‑device pipelines without writing custom C++ kernels. This democratizes edge AI in a way that mirrors the open‑source ethos of the Gemma series, which Soni emphasizes as “open‑weight models that developers can run on commodity hardware.”

Google’s move signals a shift from “AI as a cloud service” to “AI as a personal utility.” By delivering a full‑featured LLM that runs offline on a mainstream smartphone, the company not only showcases the maturity of its distillation and attention‑optimisation techniques but also nudges the industry toward a future where privacy‑preserving, low‑latency AI is the default rather than the exception. If the early benchmarks hold up, the smaller Gemma 4 variants could become the go‑to building blocks for on‑device assistants, real‑time translation, and even on‑the‑fly code generation—all without ever pinging a remote server.