PrismML Launches Ternary Bonsai 2 for On-Device AI Agents on iPhone

On October 16, AI company PrismML announced the launch of its Ternary Bonsai 2 series models in a blog post published yesterday (October 15). The series includes three model sizes: 2B, 4B, and 9B. The primary goal is to enable local execution of AI agents on smartphones, PCs, and wearable devices, supporting capabilities such as reasoning, image recognition, and tool invocation.

Built upon Alibaba’s Qwen 3.5 base model, the Ternary Bonsai 2 series utilizes "ternary" weight representation, encoding model parameters using only three values: -1, 0, and +1. According to official figures, the storage footprint for the 2B, 4B, and 9B models is approximately 0.43GB, 0.95GB, and 2GB, respectively.

In terms of capability, the Bonsai 2 series emphasizes the practicality of local agents: following instructions, completing daily tasks, invoking tools on mobile devices, and avoiding repetitive outputs. PrismML evaluated the models across six categories—intelligent tool calling, mathematics, instruction following, programming, knowledge & reasoning, and visual understanding—using 19 benchmarks. Notably, the 9B model reduces token consumption by approximately 40% for standard instruction execution and everyday mathematical tasks.

Regarding performance, PrismML states that on hardware such as the RTX 4090 and M5 Pro, the new models achieve up to 5x throughput and 3x energy efficiency compared to their full-precision counterparts. On the iPhone 18 Pro Max, the 9B model reaches a maximum speed of 34 tokens per second, while the 2B model achieves up to 120 tokens per second. IT Home has included relevant screenshots below:

This series complements the existing Bonsai 2 27B model, forming a complete product line ranging from lightweight tasks to complex workflows. Developers can run the models on NVIDIA GPUs via llama.cpp, or deploy them on Macs, iPhones, and other devices using MLX. The model weights are released under the Apache 2.0 license.