PrismML launches ternary bonsai 2 AI models with enhanced efficiency for mobile devices

PrismML announced on October 15 the release of the Ternary Bonsai 2 series, featuring models of 2B, 4B, and 9B parameters. These models are designed for local AI agent execution on mobile, PC, and wearable devices, with capabilities including reasoning, image recognition, and tool invocation. Built on Alibaba's Qwen 3.5 base, the models use a ternary weight representation with -1, 0, and +1 values, reducing storage to approximately 0.43GB, 0.95GB, and 2GB respectively. The 9B model demonstrated a 40% reduction in token usage for standard instruction execution and daily math tasks. On iPhone 18 Pro Max, the 9B model achieves up to 34 tokens per second, while the 2B model reaches 120 tokens per second. Developers can deploy the models using llama.cpp on NVIDIA GPUs or MLX on Mac, iPhone, and iPad, with weights available under the Apache 2.0 license.