Physix Frontier · News Briefing Card (IT Home · Oct 16, 2026)

PrismML launches Ternary Bonsai 2 on-device model series

KEY FACTS

PrismML released the Ternary Bonsai 2 series of models on October 15.
The series includes 2B, 4B, and 9B sizes, built on Alibaba's Qwen 3.5.
The models encode parameters with ternary weights of -1, 0, and +1.
The three models occupy roughly 0.43GB, 0.95GB, and 2GB of storage.
The 9B model cuts token usage by about 40% on instruction and math tasks.
On an iPhone 18 Pro Max, the 2B model reaches up to 120 tokens per second.

KEY DATA

0.43GB2B storage footprint

0.95GB4B storage footprint

2GB9B storage footprint

120 tokens/siPhone 18 Pro Max 2B

PHYSIX OBSERVATION

Ternary weights squeeze on-device models under 2GB, and running a 9B agent on a phone is starting to have real utility. The real signal is that local tool calling and reasoning no longer depend on the cloud, benefiting both privacy and latency. But whether the ecosystem keeps up depends on the maturity of deployment paths like llama.cpp and MLX.

Source: IT Home report