PrismML Shrinks AI Models for Personal Devices

The new additions to the Bonsai 2 family occupy just 0.43GB, 0.95GB, and 2.00GB of storage respectively. According to CEO Babak Hassibi, these specifications are nine times smaller than standard models, allowing developers to deploy agentic workflows on hardware with strict power and memory budgets. During testing on an iPhone 18 Pro Max, the 2B version achieved inference speeds of 120 tokens per second, while the 9B variant reached 34 tokens per second.

Beyond raw speed, the models demonstrate significant gains in energy efficiency. On hardware like the RTX 4090 and Macbook M5 Pro, the Bonsai 2 series delivers up to five times the throughput of full-precision alternatives. These improvements enable more complex interactions, such as interpreting visual data and executing tool-driven tasks, without tethering the device to a cloud server. The models are available now under an Apache 2.0 license, providing a framework for developers to integrate sophisticated AI assistants into consumer-grade hardware and smart glasses.