PrismML Releases Ternary Bonsai 2 Models for On-Device Agents
On October 15, AI company PrismML announced the release of the Ternary Bonsai 2 series, a family of models available in 2B, 4B, and 9B parameter sizes. The models are built on Alibaba's Qwen 3.5 base and use a ternary weight representation that encodes model parameters with only three values: -1, 0, and +1. They are designed to run AI agents locally on phones, PCs, and wearable devices, supporting reasoning, image understanding, and tool use.
The 2B, 4B, and 9B models occupy approximately 0.43GB, 0.95GB, and 2GB of storage, respectively. PrismML says the series focuses on practical local agents, including instruction following, completing everyday tasks, calling tools on mobile devices, and avoiding repetitive outputs. In evaluations across 19 benchmarks covering tool calling, math, instruction following, coding, knowledge and reasoning, and visual understanding, the 9B model reduced token usage by about 40% on standard instruction-following and everyday math tasks.
On RTX 4090 and M5 Pro hardware, the new models deliver up to 5x higher throughput and 3x better energy efficiency compared with corresponding full-precision models. On iPhone 18 Pro Max, the 9B model reaches up to 34 tokens per second, while the 2B model hits up to 120 tokens per second. The series can be combined with the existing Bonsai 2 27B model to form a complete product line from lightweight tasks to complex workflows.
Developers can run the models via llama.cpp on NVIDIA GPUs or deploy them using MLX on Mac, iPhone, and iPad. Model weights are released under the Apache 2.0 license.
