PrismML Releases Ternary Bonsai 2 AI Models for Local Edge Devices
PrismML has released the Ternary Bonsai 2 model series in 2B, 4B, and 9B sizes, built on the Qwen architecture and utilizing a ternary weight representation of -1, 0, and +1. The models achieve high local throughput, reaching up to 120 tokens per second for the 2B version on an iPhone 18 Pro Max. This release marks a significant step forward for edge AI, enabling efficient local execution of complex AI agents and tool-calling tasks on consumer smartphones and PCs with drastically reduced memory footprints and high energy efficiency. By lowering hardware resource barriers, it brings sophisticated AI capabilities closer to everyday mobile and wearable devices. The storage footprints for the 2B, 4B, and 9B models are approximately 0.43GB, 0.95GB, and 2GB respectively, and the models are available under an open Apache 2.0 license. Developers can run them via llama.cpp on NVIDIA GPUs or deploy them using MLX on Apple Silicon devices like Mac, iPhone, and iPad.
## BACKGROUND
Ternary quantization, often associated with 1.58-bit models, restricts neural network weights to three discrete values: -1, 0, and +1. This technique dramatically compresses model sizes and replaces resource-heavy multiplication operations with simpler additions and subtractions during inference, greatly accelerating performance on standard hardware.