AI/TLDR

leonickson1 · 2026-08-03 · notable

Swiftlet — 80B Qwen runs in 4.3 GB on a Mac, 35B on an iPhone

Swiftlet is an Apache-2.0 Swift and Metal runtime that streams routed MoE experts from disk on demand. It runs Qwen3-Next and Qwen3.5/3.6 hybrid MoE models on Apple hardware — 80B in 4.3 GB peak RAM on a Mac, 35B in about 2.5 GB on an iPhone 17.

Swiftlet GitHub repository social card

Swift and Metal runtime that runs Qwen 35B on an iPhone and 80B on a Mac by streaming MoE experts from local storage.

What is it?

Swiftlet is an Apache-2.0 project from developer leonickson1 that hosts Qwen3-Next and Qwen3.5/3.6 hybrid MoE models on Apple devices. The trick: only a small dense core sits in RAM, and the routed mixture-of-experts weights stream from local storage as tokens need them.

How does it work?

Every token routes through a small subset of experts, so Swiftlet only needs the active experts in memory. The routed weights stream from disk on demand while a small dense core stays resident — Swiftlet drops an 80B footprint from tens of gigabytes of RAM to about 4.3 GB peak, at the cost of extra storage I/O.

Why does it matter?

MoE models of this size have needed workstations or servers. Swiftlet makes a 35B model feasible on an iPhone 17 at about 2.5 GB peak RAM and ~1 token/second, and 80B feasible on a consumer Mac at 4.3 GB RAM and 4.5–5 tokens/second decode. No cloud call, no data leaves the device.

Who is it for?

iOS and macOS developers, on-device AI hobbyists

Try it

git clone https://github.com/leonickson1/Swiftlet

Sources · 2 outlets

Tags

  • swiftlet
  • apple
  • metal
  • swift
  • on-device
  • qwen
  • moe
  • mixture-of-experts
  • streaming
  • iphone
  • mac
  • local-llm
  • apache-2
  • repo

← All releases · Learn AI