Home›Crypto›Market Structure›PrismML’s Bonsai 27B brings a 27B reasoning model to i…

PrismML’s Bonsai 27B brings a 27B reasoning model to iPhone

The 27-billion-parameter model is compressed to 3.9 GB and runs at about 11 tokens per second on an iPhone 17 Pro Max.

Reported · Written up by our desk · Market Structure

PrismML has released Bonsai 27B, a 27-billion-parameter reasoning AI model compressed to about 3.9 GB, making it small enough to run on a consumer smartphone, according to Decrypt.

The company said the binary variant runs at roughly 11 tokens per second on an iPhone 17 Pro Max, and it is available for free under the Apache 2.0 license.

Decrypt also reported that a ternary variant, at about 5.9 GB, retains 94.6% of full-precision benchmark performance and reaches around 26 tokens per second on an M5 Pro laptop.

The model uses an end-to-end compression approach, reducing each weight from 16-bit floating point to a single sign in the binary build or three states in the ternary build, and Decrypt said Apple is in early talks with PrismML about the underlying compression technology.

More like this

Sources