September 21, 2026
A 27B-parameter model runs on a phone without internet: twice as fast on iPhone 18 Pro as on 17 Pro
Bonsai 27B’s one-bit build takes up 3,9 GB and runs on an iPhone offline: twice as fast on the 18 Pro as on the 17 Pro.
Developer Adrien Grondin ran a 27B-parameter model on an iPhone 18 Pro and recorded it answering without an internet connection.
A large model no longer needs the cloud: it fits in a phone’s memory and runs even in airplane mode.
Why it fits. In its standard form, Bonsai 27B takes 54 GB of memory. A build compressed to one bit per weight takes 3,9 GB, 14,2 times less than the original.
What changed in the hardware. The same model previously ran on an iPhone 17 Pro. Apple has now put two 16-core Neural Engine blocks and 12 GB of LPDDR5X memory with 115,2 GB/s bandwidth into the A20 Pro, 50% more than before. Grondin showed responses twice as fast.
The Bonsai authors claim 11 tokens per second on an iPhone 17 Pro Max: that is the speed of unhurried reading aloud.
How to replicate it. The test ran in the free Locally AI app, which added Bonsai 27B in version 1.62.0 dated 14.07.2026. You need iOS 18.1 or newer. For people working in a Mac terminal, the model authors provide two mlx-lm commands.
The cost of compression. The one-bit build retains 89,5% of the original model’s quality on average across 15 benchmarks, with a steeper drop on code. On the base iPhone 17 with 8 GB of memory, an independent analysis measured 9,0 tokens per second, while speed fell by 20–45% under sustained load.
iOS gives a single app roughly half of the device’s memory, about 6 GB on a 12 GB phone. The more accurate ternary build weighs 5,9–7,05 GB and no longer fits on a phone.
The ternary build is five percentage points more accurate than the one-bit version and awaits phones with more memory headroom.
