Language model peak
3.1 GB is the approximate peak resident memory of the Gemma 4 E2B language model during on-device inference.
Measured memory
What the models actually use on iPhone.
All memory figures below were measured during on-device crash forensics on TestFlight build 44, July 2026. Each entry states what the number covers.
Image input overhead
0.8 GB is the approximate additional memory consumed when an image input is attached on top of the running language model.
LIVE mode peak
4.3 GB is the approximate whole-device peak memory in LIVE mode, where the language model, voice pipeline, and camera pipeline run together.
App memory ceiling
4 GB is the approximate usable app memory ceiling on 6 GB RAM iPhones, which is why PhoneClaw gates local runtimes on available memory headroom.
Context window
4096 tokens is the conservative context window PhoneClaw currently favors, leaving room for model weights, KV cache, image or audio input, tool-call state, and iOS memory pressure. Documented in Docs/IOS_MEMORY_LIMITS.md.
| Model | Min device | Peak memory | Notes |
|---|---|---|---|
| Gemma 4 E2B (LiteRT) | A16 and above (documented, README) | ~3.1 GB resident (measured, TestFlight build 44 forensics, 2026-07) | ~2.4 GB download; the recommended default model. |
| Gemma 4 E4B (LiteRT) | iPhone 15 Pro and above (documented, README) | No measured on-device figure published yet. | ~3.4 GB download; CPU-only under the sideload-signing memory cap. |
| MiniCPM-V 4.6 | A17 Pro and above recommended (documented, README) | Image input adds ~0.8 GB on top of the language model (measured, TestFlight build 44 forensics, 2026-07) | Image Q&A and LIVE camera understanding. |
Documented requirements
Devices and download sizes, straight from the README.
These are the published requirements in the PhoneClaw repository README.
A16 and above
2.4 GB is the approximate download size of the E2B model file (gemma-4-E2B-it.litertlm). Documented in the README model preparation section.
iPhone 15 Pro and above
3.4 GB is the approximate download size of the E4B model file (gemma-4-E4B-it.litertlm). Documented in the README model preparation section.
A17 Pro and above recommended
Documented in the README model recommendation table for image Q&A and real-time camera in LIVE mode.
Port 18080
18080 is the default port of the PhoneClaw Mac Gateway, which advertises the _phoneclaw-llm._tcp Bonjour service and must run as PhoneClawGateway.app. Documented in README section 7.
Backends and throughput
Backend selection, signing caps, and what is not published.
This page only carries numbers that were measured on-device or written in repository documentation.
CPU / GPU backend selection
The inference backend (GPU or CPU) is selectable directly from the settings page; CPU is the default in sideloaded builds to fit within Sideloadly-signed memory limits. Documented in the README v1.2.2 release notes.
Sideload-signing memory cap
Under the sideload-signing memory cap, the E4B model only works on CPU (GPU will fail), and E2B is the recommended model; Xcode-signed builds are not subject to this cap and can run E2B / E4B with GPU enabled. Documented in the README v1.2.2 release notes.
Throughput
No on-iPhone tokens-per-second figure has been published yet. On-device decoding speed varies with device generation, backend (CPU or GPU), and thermal state. When measured on-device figures exist, they will appear here with the exact device and build they were measured on.