ChainThink report, April 1, 2026: According to 1M AI News monitoring, PrismML, an AI lab co-founded by mathematician Babak Hassibi from Caltech, has emerged from stealth mode and open-sourced the 1-bit Bonsai series of large language models. The flagship model, 1-bit Bonsai 8B, features 8.2 billion parameters with a memory footprint of just 1.15GB—compressed approximately 14-fold compared to comparable 16-bit models—and simultaneously released two smaller variants: 4B (0.5GB) and 1.7B (0.24GB).

The Bonsai 8B is a true end-to-end 1-bit model, where all layers—including embedding, attention, MLP, and output head—are represented exclusively using +1 or -1 weight values, without any high-precision patches. PrismML claims its inference and language understanding capabilities on standard benchmarks match those of 16-bit full-precision models. The core compression mathematics were developed over several years at Caltech by the team, with intellectual property rights held by Caltech, while PrismML holds the exclusive licensing rights. The model was trained using Google’s v4 TPU.

In benchmarked performance, the model achieves 136 tok/s on M4 Pro Mac, 440 tok/s on RTX 4090, and approximately 44 tok/s on iPhone 17 Pro Max. Standard 16-bit 8B models cannot fit into any iPhone device, and power consumption is reduced by roughly 4–5 times compared to 16-bit counterparts. PrismML notes that current hardware is not optimized for 1-bit inference; the gains in speed and energy efficiency primarily stem from drastically reduced memory footprint. If future hardware is specifically designed for 1-bit inference, efficiency could improve by another order of magnitude.

PrismML has completed a $16.25 million SAFE and seed round, backed by Khosla Ventures, Cerberus Capital, and Caltech. Vinod Khosla, founder of Khosla Ventures, commented on the achievement: “This is not a minor iteration—it’s a major technological breakthrough, a mathematical leap, not just another small model.”