PrismML has released Bonsai 27B, a 27-billion-parameter AI model that the company says is the first major model of its size capable of running entirely on a smartphone. The release marks a significant milestone in bringing advanced AI capabilities to mobile devices without requiring cloud connectivity.
The model, which runs on iPhones, iPads, and Macs, demonstrates that powerful language models can operate locally on consumer hardware. PrismML claims Bonsai 27B achieves performance competitive with much larger cloud-based models while requiring no internet connection for inference.
“This is a breakthrough for on-device AI,” said a PrismML representative. “Users can now run full reasoning AI on their phones for free, with complete privacy since no data leaves the device.”
The implications of on-device AI are significant. Running models locally eliminates latency issues associated with cloud-based AI, enhances privacy since user data never leaves the device, and reduces the cost of AI inference for both users and service providers. It also enables AI functionality in areas with limited or no internet connectivity.
Bonsai 27B uses optimization techniques including quantization and model pruning to fit within the memory constraints of mobile devices. Modern flagship smartphones typically have between 8GB and 16GB of RAM, and the model is designed to operate within those limits while maintaining acceptable inference speeds.
The release comes as major tech companies race to bring more capable AI to edge devices. Apple has been developing its own on-device AI capabilities through Apple Intelligence, while Google has optimized versions of its Gemini models for Pixel phones. Qualcomm and MediaTek have also introduced processors with dedicated AI accelerators designed for on-device inference.
PrismML’s Bonsai 27B is available for free, reflecting the company’s strategy of democratizing access to advanced AI. The model can be downloaded and used offline, making it one of the most accessible large language models for consumer use.
This article was adapted from Decrypt. Read the original here.
