Gemma 4 on Phone: Run It on iPhone or Android

Every other Gemma 4 guide on this site assumes you're sitting at a laptop or a desktop GPU. This one doesn't. If you've heard that Gemma 4 "runs on your phone" and want to know whether that means your phone, and how to actually do it without digging through GitHub issues, this is that guide.

Quick Answer

Yes — but only two of the five models are built for it. Gemma 4 ships as five sizes (E2B, E4B, 12B, 26B A4B, 31B), and only the two smallest — E2B and E4B — are edge-optimized for phones. The 12B, 26B A4B, and 31B models are not phone models; they need 16 GB+ of unified memory or dedicated GPU VRAM, which no phone has. See our hardware requirements guide for those four.

• E2B needs roughly 6 GB of RAM — comfortable on an iPhone 13 Pro or newer, or almost any Android phone from the last four years.
• E4B needs roughly 8 GB of RAM — an iPhone 15 Pro/16 Pro or newer, or a recent Android flagship.
• Both are natively multimodal (text, vision, audio) and both run fully offline once downloaded.

Which iPhones and Android Phones Qualify

RAM is the gating factor, not chip speed — a model either fits in memory or it doesn't. Here's how the last few generations stack up.

iPhone

Model	RAM	E2B (6 GB)	E4B (8 GB)
iPhone 13 Pro / Pro Max	6 GB	Yes	No
iPhone 14 Pro / Pro Max	6 GB	Yes	No
iPhone 15 / 15 Plus	6 GB	Yes	No
iPhone 15 Pro / Pro Max	8 GB	Yes	Yes
iPhone 16 / 16 Plus	8 GB	Yes	Yes
iPhone 16 Pro / Pro Max	8 GB	Yes	Yes

Figures are confirmed device specs (GSMArena, TechRadar, MacRumors). Standard iPhone 13/14 (non-Pro) and anything with less than 6 GB isn't a safe bet for E2B — check your exact model's RAM if you're on an older or entry-level device.

Android

Android RAM varies more by SKU than iPhone does, so treat this as a rough guide rather than a per-model guarantee:

Tier	Typical RAM	E2B (6 GB)	E4B (8 GB)
Budget/entry (2022+)	4–6 GB	Maybe (tight)	No
Mid-range (Pixel 8, Galaxy A-series)	8 GB	Yes	Yes
Flagship (Galaxy S24/S25, Pixel 9)	8–12 GB	Yes	Yes
Flagship Ultra/Pro (Galaxy Ultra, Pixel Pro, OnePlus)	12–16 GB	Yes	Yes

If you don't know your phone's RAM, check its spec sheet before downloading — under 6 GB total system RAM won't leave enough free RAM for E2B once the OS and other apps take their share.

Prefer an instant answer? Our interactive Can You Run Gemma 4? checker now has a Phone / Tablet tab — pick your platform and RAM, get a verdict in seconds. It covers Mac, PC, and CPU-only setups too.

3 Ways to Run Gemma 4 on Your Phone

1. Google AI Edge Gallery (official, zero setup)

This is Google's own app and the place most people should start. It's confirmed live on the App Store and Google Play, open-source (Kotlin on Android, Swift on iOS) via the google-ai-edge/gallery repo, and requires Android 12+ or iOS 17+.

1. Install Google AI Edge Gallery from the App Store or Play Store (a sideloadable APK is also on the GitHub releases page).
2. Open the app and browse the model catalog — Gemma 4 E2B and E4B are listed as featured/official models.
3. Tap to download over Wi-Fi.
4. Once downloaded, everything runs on-device. Start a chat, drop in a photo, or try the built-in Thinking Mode to watch the model's reasoning steps.
5. No account, no API key, no internet after the initial download — Google's own description calls it "100% On-Device Privacy."

Google's developer blog notes E2B can run in under 1.5 GB of RAM on some devices, thanks to aggressive 2-bit/4-bit quantization plus the same Per-Layer Embeddings trick used on desktop models (see our hardware requirements guide). That's runtime memory, not download size — Google hasn't published an exact in-app download-size table, so treat any specific GB figure for the Gallery app as an estimate until you've checked it yourself.

2. Alternative apps (PocketPal, Off Grid, and other llama.cpp-based clients)

If you'd rather use a general-purpose local-LLM app, several llama.cpp-based mobile apps work with Gemma 4 in GGUF format:

• PocketPal AI (iOS, Android) — open source, built on llama.cpp, and lets you search Hugging Face directly from the app. Gemma 4 E2B/E4B GGUF quants are published there (see our download guide for the exact repos), so you can search by name and pick a quant that fits your memory.
• Off Grid — another open-source, offline-first app (llama.cpp/llama.rn) that community reviewers have specifically used to run Gemma 4's edge models. Community write-ups on dev.to report pulling the Q4_K_M GGUF quant of E2B (roughly 1.3 GB) and E4B (roughly 2.5 GB) — smaller than the "5 GB" Q4_0 figure in our desktop hardware guide because it's a different quant scheme built for phones, not GPUs. More on that below.

Both apps let you browse and swap quantization levels if a download feels too large for your storage.

3. For developers: MediaPipe LLM Inference API / LiteRT

If you're building your own app, Google's on-device stack is the MediaPipe LLM Inference API, layered on LiteRT (formerly TensorFlow Lite) and the newer LiteRT-LM runtime — available for Android, iOS, and Web, and what the AI Edge Gallery app itself is built on.

Worth knowing: Google's docs note the classic MediaPipe LLM Inference API is now in maintenance-only mode, with new work going into LiteRT-LM — start there for a new project. Models need .task or .litertlm format, not raw safetensors or GGUF; Google publishes ready-made LiteRT builds of the edge models for this. Full setup (SDK install, conversion, code samples) is in the official docs: LLM Inference guide for Android and Deploy Gemma on mobile devices.

This route is overkill if you just want to chat with the model — use the Gallery app or PocketPal for that instead.

What to Expect: Speed, Battery, and Heat

Speed. These numbers are all community-reported (dev.to hands-on posts), not official Google benchmarks, so treat them as ballpark figures that will vary by device, background load, and which quant you pick:

Model	Reported speed	Device class
E2B	~12–18 tok/s	iPhone with Metal acceleration
E2B	~12–20 tok/s	Recent Snapdragon flagship chip
E4B	~10–15 tok/s	iPhone 15 Pro / 16 Pro
E4B	~8–15 tok/s	Recent Snapdragon flagship chip

Google's own published numbers are for different hardware entirely — 133 prefill/7.6 decode tok/s on a Raspberry Pi 5 CPU, and 3,700 prefill/31 decode tok/s on a Qualcomm Dragonwing IQ8 with NPU acceleration (source) — a sanity check that the architecture scales, not a substitute for phone-specific numbers.

Battery and heat. Be honest with yourself here: running any LLM locally is sustained, heavy compute — closer to gaming or video export than typical app use. Expect your phone to warm up on longer sessions and battery to drain faster than normal; this is true of any on-device model, not something specific to Gemma 4. Short bursts are fine; long back-and-forth conversations will cost noticeably more battery than the same chat with a cloud chatbot.

One optimization: several community guides mention switching KV cache to q4_0 in an app's settings, which can meaningfully speed up longer conversations at a small quality cost.

A Note on File Sizes: Why the Numbers Don't Match

Our desktop hardware requirements guide lists E4B at 5 GB (Q4_0). This page cites community reports closer to 2.5 GB. Both can be right — they're not measuring the same thing:

• 5 GB Q4_0 — Google's official minimum GPU/TPU memory to load the desktop-format weights (Ollama, LM Studio, workstation GPU).
• ~2.5 GB — download size for a phone-tuned GGUF Q4_K_M quant used by llama.cpp apps (PocketPal, Off Grid): a different quantization scheme, packed for storage-constrained phones rather than VRAM.
• Google's in-app LiteRT/.litertlm builds (AI Edge Gallery) are a third format, with Google citing under 1.5 GB of runtime RAM for E2B rather than a download size.

Same model, three packagings. If a number you see elsewhere doesn't match this site, check which format it's describing before assuming one is wrong.

Limitations

• Context length is shorter in practice than the spec sheet. E2B and E4B officially support up to 128K tokens, but that's a desktop/GPU-memory number. On a phone, expect comfortable use in the low thousands of tokens before RAM pressure slows things down, not the full 128K.
• No 12B, 26B A4B, or 31B on phones. These need 16 GB+ of memory no current phone has. For that quality on the go, use a cloud API or a remote server you control from your phone (Ollama's REST API — see our Ollama guide).
• Audio and video support is uneven across apps. The models handle audio/video architecturally, but whether a given app exposes that in its UI varies — check the app's release notes rather than assuming it works everywhere.
• Thermal throttling is real. Sustained generation on a hot phone slows down over time, not just drains the battery.

For a broader look at what E2B and E4B are good for beyond mobile (IoT, embedded devices, and more), see our features and use cases guide.

FAQ

Can iPhone run Gemma 4? Yes, if it's an iPhone 13 Pro or newer (6 GB+ RAM) for E2B, or an iPhone 15 Pro/16 Pro or newer (8 GB+ RAM) for E4B. Use the official Google AI Edge Gallery app, which requires iOS 17 or later.

Which is better on a phone, E2B or E4B? E4B gives noticeably better reasoning and output quality — it's the larger of the two edge models — but it needs 8 GB of RAM and runs somewhat slower in community benchmarks (roughly 8–15 tok/s vs. E2B's 12–20 tok/s). If your phone qualifies for both, E4B is worth trying first; drop to E2B if it feels sluggish or your phone struggles with the memory footprint.

Does it work offline? Yes — that's the entire point of running Gemma 4 on-device. After the initial model download (which does require an internet connection), every app covered here — Google AI Edge Gallery, PocketPal, Off Grid — runs completely offline, with no data leaving your phone.

Why can't my phone run the 12B model? It needs about 16 GB of unified memory or GPU VRAM at Q4 (see our hardware requirements guide for exact numbers), and no phone ships with that much RAM available to a single app. E2B and E4B exist because Google built them with edge-specific tricks — Per-Layer Embeddings and Hybrid Sliding Window Attention — that 12B, 26B A4B, and 31B skip, since those three target laptops, desktops, and workstations instead.