Gemma 4's 2B Edge Model Runs Fully Offline on iPhone — 1.5 GB, Apple Neural Engine, No Cloud Required
Google’s Gemma 4 release included more than the headline 31B model. The edge variants — E2B (2 billion parameters) and E4B (4 billion parameters) — are now running fully offline on iPhones, using Apple’s Neural Engine as the inference backend. The path to deployment is straightforward: download the quantized model (~1.5 GB) through either Locally AI or Google’s own AI Edge Gallery app, and all subsequent processing happens on-device.
No network call. No API key. No cloud dependency.
What “Edge-Optimised” Means in Practice
Google engineered the E2B and E4B models specifically to extract more capability per parameter than previous-generation Gemma variants. The result is documented in third-party testing: on standard benchmark tasks, the 2.3B Gemma 4 E2B model performs comparably to what 70B models produced 18 months ago, at roughly 1.5 GB of RAM. On an iPhone 15 Pro or later, where the Apple Neural Engine provides substantial dedicated matrix multiplication throughput, inference runs without throttling the rest of the device.
For reference, Apple Intelligence models — Apple’s own on-device AI layer — operate in a similar parameter range and use the same neural engine. Gemma 4 E-series is essentially slot-compatible with that execution environment.
Deployment Paths
Two consumer apps currently support the Gemma 4 edge models on iOS:
- Locally AI — a local inference app that supports multiple quantised open models; Gemma 4 E2B and E4B are available as direct downloads from within the app
- Google AI Edge Gallery — Google’s own first-party showcase app for on-device model inference; added Gemma 4 support alongside the model’s April 2026 release
Both apps handle model management and quantisation automatically. The underlying runtime in both cases maps the model’s operations onto Apple’s Neural Engine acceleration stack.
The Practical Implication
Gemma 4 E2B running on consumer hardware without a network dependency means two things:
Privacy-first AI becomes a commodity option. For sensitive query types — health, finance, legal, personal communications — an on-device inference path that requires no data to leave the device is now available to anyone with a mid-range iPhone and a 1.5 GB download budget. The capability exists; the friction to access it is minimal.
Cloud providers lose a forcing function. A meaningful fraction of LLM use cases — summarisation, drafting, Q&A on local documents — do not require frontier-scale intelligence. If Gemma 4 E2B covers those cases at zero marginal cost and without latency, the justification for API calls weakens on purely economic grounds.
Google released all Gemma 4 variants under Apache 2.0, which means the edge models can be embedded in commercial products without licence restrictions or royalty obligations. The iOS deployment is a consumer-facing signal of what that licence structure enables at scale.