// Device Check

Is your iPhone
AI-ready?

Most people have never run an AI model on their phone — because nobody told them it was possible. Pick your device and find out what it can do, completely offline.

> No guide like this exists anywhere else. We built it because we had to.

All on-device AI is 100% private — nothing is ever sent to a server.

// Why models sometimes crash

Silent crashes — explained

Most on-device AI apps don't tell you why the app closes. Here's what's actually happening, and how Mulberry handles it for you.

// Cause 01
Model too large for your device
iOS enforces a per-app memory budget (jetsam). A 4.5 GB model on a 4 GB iPhone hits that ceiling instantly — the OS kills the app silently, no error shown.
→ Fix: download only models in your tier
// Cause 02
KV cache growth mid-conversation
Even models that fit at turn 1 can crash by turn 3–4. Each reply grows the key-value cache in memory. 7B-class models on 6 GB devices are especially vulnerable.
→ Fix: Mulberry caps the KV cache automatically
// Cause 03
Thinking mode on Gemma 4 E4B
Gemma 4 E4B supports a reasoning / "thinking" mode that expands its KV cache exponentially. This crashes every iPhone regardless of RAM — it's a model architecture issue, not a device issue.
→ Fix: Mulberry disables thinking mode on E4B
// Cause 04
Wrong model for your chip
Google's "edge" label on Gemma 4 E4B means edge vs. datacenter — not edge vs. your iPhone. The E4B needs 4 GB free, so it can't run on iPhone XR or 11 class hardware.
→ Fix: use the device selector to see your options

// The three tiers

What your hardware unlocks

Lite
iPhone XR / XS / 11 series · iPad 8th–9th gen · A12 / A13 · ~3 GB RAM
Quick facts, short replies, simple code under 15 lines — all offline. Accuracy softens beyond a paragraph. The new Gemma 4 E2B is a significant quality upgrade over the older Gemma 2 2B at the same size.
Gemma 4 E2B · 1.5 GB ★ New Gemma 2 2B · 1.2 GB Llama 3.2 1B · 0.7 GB SmolLM2 1.7B · ~1 GB Qwen3 0.6B · 0.4 GB
Core
iPhone 12 / 13 / 13 mini · iPad Air 4+ · A14 / A15 · 4 GB RAM
The everyday workhorse. Real code generation, full-length documents, short stories up to ~800 words, and debugging — correctly, first try. Phi-4 mini is a direct upgrade from Phi-3.5 with the same footprint. Note: Gemma 4 E4B runs at this tier but thinking mode must stay off — enabling it crashes every iPhone regardless of RAM.
Phi-4 mini · 2.2 GB ★ New Phi-3.5 mini · 2.2 GB Llama 3.2 3B · 2 GB Qwen 2.5 3B · 2 GB Qwen3 1.7B · ~1.2 GB Gemma 4 E4B · 2.8 GB ⚠ no thinking mode
Magnum Brain Tuning
iPhone 13 Pro+ · 14 / 14 Plus and newer · iPad Pro / Air M-series · 6–12 GB RAM
Production-quality output first try. 7B-class models unlock at this tier. On 6 GB devices (13 Pro, 14, 15) the 7B models work but may crash after 3–4 long turns — keep sessions short or use a 4B model instead. iPhone 15 Pro and newer (8 GB+) handle 7B models indefinitely.
Qwen3 4B · 2.8 GB ★ New Gemma 3 4B · ~2.5 GB SmolLM3 3B · ~2 GB Qwen 2.5 7B · 4.8 GB Mistral 7B · 4.5 GB 7B on 6 GB: short sessions only
lock

Whatever tier your device runs, the AI lives on your phone. No cloud round-trip, no account, no telemetry. Your prompts and your work never leave the device.