PrismML launches Ternary Bonsai 2 models to run AI agents on phones

AI company PrismML has launched the Ternary Bonsai 2 model family in three sizes: 2B, 4B, and 9B. The goal is to run AI agents locally on phones, PCs, and wearables, handling reasoning, image recognition, and tool calls.

The models are built on Alibaba's Qwen 3.5 base and use a ternary weight method, encoding parameters with only -1, 0, and +1. This keeps storage tiny: about 0.43GB, 0.95GB, and 2GB for the three sizes.

The 9B model needs about 40% fewer tokens for standard instructions and everyday math. On hardware like the RTX 4090, the new models deliver up to 5x throughput and 3x energy efficiency over full-precision versions. On the iPhone 18 Pro Max, the 9B model hits 34 tokens per second.

Developers can run the models with llama.cpp on NVIDIA GPUs or MLX on Macs and iPhones. The weights are open under Apache 2.0 and pair with the existing Bonsai 2 27B for a full product line.