August 3, 2026
Pocket AI, peak chaos
Show HN: Run an 80B Qwen in 4.3 GB of RAM on a Mac, and a 35B on an iPhone
Your iPhone Can Run a Giant AI Now, but the comments are screaming about speed and dead SSDs
TLDR: A developer showed off a way to run a huge AI on a Mac with surprisingly low memory use — and even get a smaller version working on an iPhone. The community instantly split between “this is the future” and “congrats on inventing a very expensive way to torture your storage drive.”
A new Show HN post dropped a classic internet bombshell: someone built Swiftlet, a tool that can run surprisingly large AI models on ordinary Apple gear, including a Mac and even an iPhone. On paper, it sounds almost absurd — a model usually associated with heavyweight hardware now squeezed into just a few gigabytes of memory. The creator says the 35B version can even run fully on-phone with no server involved, which immediately triggered the usual mix of awe, suspicion, and gleeful nitpicking.
And wow, the comments did not disappoint. The loudest skeptics basically called it a science-fair stunt with an SSD death wish. One person joked we’re now “vibe coding NAND burners,” while another sneered that these tricks are just “slow as hell” and could “kill your drive early.” Translation for normal humans: yes, it runs, but critics think it may chew through storage and take too long to be truly useful.
But the defenders came in hot too. Their argument? This is exactly how breakthroughs start — messy, impractical, and kind of ridiculous until suddenly it isn’t. One commenter was already dreaming about the day a trillion-parameter AI runs from a cheap external drive instead of a room full of expensive Nvidia chips. Others saw this as a clue to Apple’s long game: maybe your future everyday AI really does live on your phone. The most grounded crowd just wanted a practical tweak: if their laptop has more memory, can they use it to speed this thing up? In other words, beneath the doomposting and hardware horror jokes, people could smell a real idea here.
Key Points
- •Swiftlet runs Qwen mixture-of-experts models on Apple devices by keeping dense weights in memory and streaming expert weights from storage on demand.
- •The article reports Qwen3.6-35B-A3B using 2.6 GB peak RAM on a Mac and about 2.5 GB RAM on an iPhone 17, while Qwen3-Next-80B-A3B uses 4.3 GB peak RAM on a Mac.
- •Reported Mac decode speeds are 7 to 11 tok/s for the 35B model and 4.5 to 5 tok/s for the 80B model; the iPhone run is about 1 tok/s.
- •Swiftlet uses a .qpack container, single-pread SSD expert fetches, bounded LFU-plus-recency caching, and Metal runtime-compiled shaders.
- •The article says 75% of layers use Gated DeltaNet linear attention with fixed-size recurrent state, which avoids a growing KV cache for those layers.