A 27-billion-parameter model finally fits in a phone

PrismML's Bonsai 27B model fits on an iPhone thanks to extreme compression. A major breakthrough for running local agents at no cost.

PrismML runs a 27-billion-parameter class model on a phone, something that had never been done before. Bonsai 27B, backed by Qwen3.6 27B, is released in open source under the Apache 2.0 license.

The obstacle was arithmetic. A model of this size occupies about 54 gigabytes in 16-bit precision, and even good 4-bit compression leaves a footprint of around twenty gigabytes, out of reach for both a phone and most laptops. PrismML's method consists of reducing each weight to extreme values. The ternary version allows only three values (minus one, zero, plus one), or 1.71 effective bits per weight, for a 5.9-gigabyte file, tailored for a laptop. The 1-bit version is limited to two values and drops to 3.9 gigabytes, bringing it within the memory budget of a recent iPhone.

The compression runs across the entire network, with no escape to higher precision, and both versions remain multimodal: they see screenshots, documents, and camera images, with a context of 262,000 tokens. PrismML claims to retain, across a battery of fifteen benchmarks, 95% of the original model's performance for the ternary version and 90% for the 1-bit version, specifying that mathematics, code, and tool calling are the least affected capabilities.

The stated stake is that of agents. An agent does not make one call, it makes hundreds of them, each billed and traversing the network with the user's files, screen, and data. Fitting a model capable of going the distance on the device eliminates the marginal cost of iteration and keeps the data on the machine. PrismML was born from a team out of Caltech.