PrismML has introduced the Ternary Bonsai 2 series, a collection of AI models available in 2B, 4B, and 9B parameter sizes. These models are designed for local execution on smartphones, PCs, and wearable devices, enabling tasks such as inference, image recognition, and tool invocation.
The series is built upon the Qwen 3.5 architecture and utilizes a ternary weight representation, encoding parameters using only -1, 0, and +1. This approach results in storage requirements of approximately 0.43GB for the 2B model, 0.95GB for the 4B model, and 2GB for the 9B model.
Performance metrics indicate that the models achieve up to 5 times the throughput and 3 times the energy efficiency compared to full-precision counterparts on hardware like the RTX 4090 and M5 Pro. On an iPhone 18 Pro Max, the 9B model reaches 34 tokens per second, while the 2B model achieves 120 tokens per second.
The models are compatible with the existing 27B version, creating a comprehensive product line for various workloads. Developers can deploy the models via llama.cpp on NVIDIA GPUs or through MLX for Mac, iPhone, and iPad devices. The weights are released under the Apache 2.0 license.
