PrismML Releases Ternary Bonsai 2 Models for Phones
PrismML is pushing the same quantization method it used on a 27B model down to phone-sized weights, where the 0.43GB footprint and the iPhone throughput numbers are what make a local assistant plausible.
Reporting from 1 source: GIGAZINE.
PrismML released Ternary Bonsai 2 2B, 4B, and 9B on October 15, 2026. The models are ternary versions of the Qwen3.5 base models, using the values -1, 0, and 1, with memory use of 0.43GB, 0.95GB, and 2GB. Ternary Bonsai 2 2B runs at 120 tokens per second on the iPhone 18 Pro Max, and the Locally AI app supports them.
The models are ternary versions of Qwen3.5 2B, 4B, and 9B, each modified to process using three values: -1, 0, and 1. Memory use stays under one-ninth of the base models, and PrismML says performance degradation from the base is suppressed, with some tests scoring above the base.
A demo shows Ternary Bonsai 2 9B handling agent tasks on an iPhone 18 Pro Max. The free app Locally AI already supports all three, and Adrien Grondin, its developer, posted a video of the 2B model running on an iPhone 18 Pro.
- Ternary Bonsai 2 2B: Based on Qwen3.5 2B, 0.43GB, 120 tokens per second on iPhone 18 Pro Max
- Ternary Bonsai 2 4B: Based on Qwen3.5 4B, 0.95GB
- Ternary Bonsai 2 9B: Based on Qwen3.5 9B, 2GB, demoed on iPhone 18 Pro Max handling agent tasks
Synthesized by Yomimono from the 1 cited source below, including Japanese-language reporting where cited, then editorially reviewed before publishing.