Gemma 4 on iPhone
Yes, the small ones. Gemma 4 E2B runs offline on iPhones with 8 GB of RAM or more in Pocket AI Lab, and on an iPhone Air it generated 35.0 tokens per second with a quality score of 63.6/100. E4B should fit on 12 GB iPhones by the app's own memory check, though we haven't benchmarked it yet. The 12B, 26B A4B and 31B models need more memory than any iPhone gives an app.
Measured 2026-10-23 in Pocket AI Lab 1.1.0 on iPhone Air (A19, 11.5 GB RAM as reported by iOS, iOS 26.6.1). Gemma 4 specs and file sizes checked 2026-10-28 against Google's model card and Hugging Face.
Which Gemma 4 models can run on an iPhone?
The two edge models, E2B and E4B, which Google built for phones. E2B is in the Pocket AI Lab catalog and measured. 12B, 26B A4B and 31B are made for laptops, GPUs and workstations.
| Model | Parameters | Context | Download | RAM needed | Measured on iPhone Air | Runs on |
|---|---|---|---|---|---|---|
| Gemma 4 E2B | 2.3B effective (5.1B with embeddings) | 128K | 4.1 GB | 5.0 GB | 35.0 tok/s, 63.6/100 | 8 GB and 12 GB iPhones |
| Gemma 4 E4B | 4.5B effective (8B with embeddings) | 128K | 6.0 GB | ~6.1 GB | not measured yet | 12 GB iPhones |
| Gemma 4 12B | 11.95B | 256K | 7.3 GB | ~8.4 GB | – | No iPhone |
| Gemma 4 26B A4B | 25.2B total, 3.8B active | 256K | 18.1 GB | ~18.9 GB | – | No iPhone |
| Gemma 4 31B | 30.7B dense | 256K | 19.5 GB | ~20.4 GB | – | No iPhone |
Parameters and context from Google's Gemma 4 model card. Download is unsloth's 4-bit GGUF on Hugging Face (Q4_K_M, UD-Q4_K_M for 26B A4B) plus the vision projector. RAM for E2B is the catalog budget, for the others it's the estimate the app's import check makes from the file size (~). Speed is generation tok/s, quality is the score on suite v3.
Which iPhone can run which Gemma 4?
Gemma 4 E2B (GGUF, Vision) runs on iPhones with 8 GB of RAM or more. On 8 GB phones it fits with little headroom, so long chats may make it reload. E4B fits only on the 12 GB models (iPhone 17 Pro, 17 Pro Max, iPhone Air), and no iPhone has room for 12B or anything bigger.
| RAM | iPhones | E2B | E4B | 12B, 26B A4B, 31B |
|---|---|---|---|---|
| 3–4 GB | iPhone 11, 11 Pro, 12, 12 mini, 13, 13 mini, SE (2nd and 3rd gen) | ✗ Too big | ✗ Too big | ✗ Too big |
| 6 GB | iPhone 12 Pro, 13 Pro, 14, 14 Plus, 14 Pro, 15, 15 Plus | ✗ Too big | ✗ Too big | ✗ Too big |
| 8 GB | iPhone 15 Pro, 16, 16 Plus, 16 Pro, 16e, 17, 17e | ≈ Runs, little headroom | ✗ Too big | ✗ Too big |
| 12 GB | iPhone 17 Pro, 17 Pro Max, iPhone Air | ✓ Runs | ✓ Fits (estimate) | ✗ Too big |
Checked against the GPU memory the app measures on each class of iPhone (October 2026): 2.12 GB on 3–4 GB, 3.10 GB on 6 GB, 5.34 GB on 8 GB, 8.06 GB on 12 GB. ✓ fits with room for long chats, ≈ fits with little headroom, ✗ too big. Before a download the app checks your exact phone.
How fast is Gemma 4 E2B on iPhone?
35.0 tokens per second on an iPhone Air, averaged over three 200-token answers (± 2.3), without thermal throttling. Two separate Build test runs, on 2026-10-21 and 2026-10-23, measured 34.9 and 35.0 tok/s.
On the llama-bench style probe it read a 512-token prompt at about 708 tok/s and wrote 128 tokens at 32.8 tok/s. Cold load took 2.6 s. The app's memory footprint peaked at 1,490 MB, far below the 4.1 GB of files, because llama.cpp maps the model from storage instead of copying it into memory. The whole file still has to fit in the GPU working set, which is why the catalog budgets 5.0 GB.
Other models of the same size landed close by: LFM2 2.6B (GGUF) at 35.9 tok/s (60.2/100) and Qwen 3.5 2B (GGUF, Vision) at 36.6 tok/s (64.4/100). Pocket AI Lab offers Gemma 4 E2B first on 12 GB iPhones, where it has room to spare and adds photo input, and LFM2 2.6B (GGUF) on 6 and 8 GB ones.
Quality by category
It passed 75 of 118 tasks on suite v3, graded by deterministic checks with greedy decoding.
Is Gemma 4 better than Gemma 3 on iPhone?
On our iPhone Air, yes. Gemma 4 E2B generated 35.0 tok/s against 20.7 for Gemma 3 4B on the same engine and scored 63.6 against 60.2. Gemma 3 4B also heated the phone enough to throttle during its run.
| Model | Download | tok/s | Quality | Input |
|---|---|---|---|---|
| Gemma 4 E2B (GGUF, Vision) | 4.1 GB | 35.0 | 63.6 | text, image |
| Gemma 2 2B | 1.5 GB | 30.3 | 40.7 | text |
| Gemma 3 4B (Vision) | 2.8 GB | 21.8 | 57.6 | text, image, video |
| Gemma 3 4B (GGUF, Vision) | 3.4 GB | 20.7 | 60.2 | text, image |
| Gemma 3 12B (Vision) | 7.5 GB | – | – | text, image, video |
Every Gemma model in the Pocket AI Lab catalog, newest run on suite v3, iPhone Air. Gemma 3 12B hasn't been benchmarked on the phone yet.
Why don't Gemma 4 12B, 26B A4B and 31B run on iPhone?
They don't fit in memory. The most iOS gives an app for a llama.cpp model is about 8.06 GB of GPU memory, on 12 GB iPhones, and each of these needs more than that before a chat even starts.
- Gemma 4 12B is a 7.1 GB file at Q4_K_M. Google puts its 4-bit weights alone at 6.7 GB, and with room for the context the app's estimate comes to 8.4 GB. Google made it for laptops with 16 GB of memory.
- Gemma 4 26B A4B uses only 3.8B parameters per token, which makes it quick on a computer, but all 25.2B have to sit in memory. That's 14.4 GB at 4-bit by Google's own table, more than the total RAM of any iPhone.
- Gemma 4 31B is a dense model and needs about 17.5 GB at 4-bit for the weights alone.
On a phone, E2B is the Gemma 4 to use. The large ones belong on a Mac or a PC with a big GPU.
Can Gemma 4 read photos on iPhone?
Yes. Gemma 4 E2B in Pocket AI Lab takes images as well as text, so you can attach a photo or take one with the camera and ask about it, with no connection.
Vision works through a 990 MB projector file (mmproj) that downloads together with the model. Google's model also understands audio, but the app doesn't pass sound to Gemma 4. Voice messages are turned into text on the phone by Whisper, once you've installed a Whisper model, and reach Gemma 4 as text. The catalog runs E2B with an 8,192-token context, far below the 128K the model supports.
How do I run Gemma 4 on my iPhone?
Install Pocket AI Lab, download Gemma 4 E2B from the Models tab and start a chat. It's one 4.1 GB download, and after that it works in airplane mode.
Install Pocket AI Lab
Get it from the App Store. It's free, 41.1 MB, needs iOS 26.1 and doesn't ask for an account.
Open Gemma 4 E2B
In the Models tab tap Gemma 4 E2B (GGUF, Vision). The app has already checked your phone's memory and tells you whether it fits.
Download it once
4.1 GB from Hugging Face, the 3.1 GB model plus the 990 MB vision projector. Wi-Fi is best, and the download keeps going in the background.
Chat offline
Start a chat with it and type, or attach a photo. Turn on airplane mode and it keeps answering.
On a 12 GB iPhone you can try E4B as well. Tap ⊕ in the Models tab and paste unsloth/gemma-4-E4B-it-GGUF. The app preselects Q4_K_M and shows its RAM and disk verdict before anything downloads. We haven't run E4B through the benchmark yet, so there's no speed figure for it here.
Questions
Can an iPhone run Gemma 4?
Yes, the small ones. Gemma 4 E2B runs offline on iPhones with 8 GB of RAM or more in Pocket AI Lab, and on an iPhone Air it generated 35.0 tokens per second with a quality score of 63.6/100. E4B should fit on 12 GB iPhones by the app's own memory check, though we haven't benchmarked it yet. The 12B, 26B A4B and 31B models need more memory than any iPhone gives an app.
Which Gemma 4 model should I use on iPhone?
Gemma 4 E2B. Pocket AI Lab recommends it first on 12 GB iPhones, and it also runs on 8 GB iPhones with little headroom. E4B is worth a try on a 12 GB iPhone. 12B and the bigger models don't fit.
How fast is Gemma 4 on iPhone?
Gemma 4 E2B generated 35.0 tokens per second on an iPhone Air (A19) in Pocket AI Lab 1.1.0, which is faster than most people read. iPhones with older chips are slower.
Does Gemma 4 work offline on iPhone?
Yes. After a one-time 4.1 GB download, Gemma 4 E2B runs entirely on the phone, airplane mode included, and conversations stay on the device.
How much storage does Gemma 4 need on iPhone?
4.1 GB for Gemma 4 E2B: the 3.1 GB model file and the 990 MB vision projector. E4B at Q4_K_M with its projector takes about 6.0 GB.
Can a 6 GB iPhone run Gemma 4?
Not in Pocket AI Lab. The GGUF build needs more than the 3.1 GB of GPU memory iOS gives an app on 6 GB iPhones. Google AI Edge Gallery runs Gemma 4 on Google's LiteRT-LM runtime, which Google lists with a smaller memory footprint. In Pocket AI Lab the pick for 6 GB is LFM2 2.6B (GGUF).
Can Gemma 4 describe photos on iPhone?
Yes. Gemma 4 E2B in Pocket AI Lab takes images as well as text, so you can attach a photo or take one with the camera and ask about it, with no connection.
Is Gemma 4 free?
Yes. Google releases Gemma 4 under the Apache 2.0 license, which allows commercial use. Pocket AI Lab is free too, with its source code under the MIT license.
More models for every iPhone are in the best local LLM for iPhone and the catalog. New to offline AI? Read how to run an LLM offline on iPhone.