AI/ML●●Solid
Running Gemma 4 on an iPhone 13 Pro
Clean Swift wrapper for Gemma 4 with vision and audio on iPhone.
Niche GemShip It
dengjiuhong
102mo ago
Back to browse
GitHub Repository
Neutral, reproducible benchmark for local LLMs on Apple Silicon (Mac · iPhone · iPad) — MLX, llama.cpp, CoreML, Apple Foundation Models
29 starsSwift
iPhone ANE holds LLM tok/s while MLX and LiteRT thermal-throttle
by mlboy·Jun 4, 2026·1 point·0 comments
AI Analysis
●●●BangerDark HorseBig BrainSolve My Problem
LiteRT beats MLX on Gemma memory while CoreML sips power on the Neural Engine.
Strengths
- •Automated `devicectl` headless mode removes manual testing friction on iOS devices.
- •Compares Google LiteRT against Apple MLX and CoreML on mobile hardware.
- •Reveals Neural Engine memory efficiency versus GPU throughput tradeoffs clearly.
Weaknesses
- •"iPhone 17 Pro" label raises eyebrows since the device doesn't publicly exist.
- •Limited model coverage favors Gemma and Qwen, needs broader architecture testing.
Category
Target Audience
iOS AI developers, Edge ML engineers
Similar To
MLC Bench · Llama.cpp Benchmarks · Perfetto
Similar Projects
AI/ML●●●Banger
A tiny C program where an LLM rewires its DAG while running
LLM mutates the workflow DAG mid-run via a constrained four-verb grammar.
Big BrainWizardryZero to One
mrkn1
1541mo ago
Developer Tools●●Solid
Slopsome – a VRAM fit calculator and tok/s database for local LLMs
VRAM calculator with crowd-sourced tok/s benchmarks when model cards already exist.
Niche GemSolve My Problem
NexAIGuy
307d ago
AI/ML●●Solid
I ran Qwen3.5 35B on my iPhone at 5.6 tok/SEC
Runs 19.5GB Qwen3.5 on 12GB RAM iPhone via memory swapping.
WizardryBold Bet
alexintosh
423mo ago
AI/ML●●●Banger
We've built a standalone Apple Watch app running LLMs offline, locally
First standalone Apple Watch LLM app running 700M-1.7B models completely offline.
WizardryZero to OneDark Horse
pielouNW
3017d ago
AI/ML●●●Banger
I wrote an LLM inference engine in pure Go – 48 tok/s zero dependencies
Pure Go LLM inference, zero dependencies, 48 tok/s—genuinely novel for Go ecosystem.
Zero to OneWizardryBig Brain
computerex
203mo ago