PrismML

PrismML (legally Prism ML, Inc.) is an AI lab founded by a group of Caltech researchers and led by co-founder and CEO Babak Hassibi, a Caltech professor whose work centers on compression. Its other co-founders are Sahin Lale and Omead Pooladzandi, who co-lead research, and Reza Sadri, and its advisers include Databricks co-founder Ion Stoica. The company says it was founded with support from Khosla Ventures, Cerberus and Google, with continuing support from Samsung, and it has raised a seed round of about 22 million dollars.

PrismML does not train frontier models from scratch. It takes existing open models and compresses their weights so aggressively that they run on laptops, desktop graphics cards and phones, a family it calls Bonsai. The first Bonsai models, at 8 billion, 4 billion and 1.7 billion parameters, arrived in March 2026, followed in July by a 27 billion parameter model in a ternary version for laptops and a one-bit version for recent iPhones. In September 2026 it released Ternary Bonsai 2 27B, a compression of Alibaba's Qwen3.8 27B that stores each weight as minus one, zero or one and fits in 5.9 gigabytes. PrismML reports it keeps 98.2 percent of the original model's aggregate benchmark score, up from about 95 percent for its first ternary 27B model. The weights are released under Apache 2.0.

Its bet is that the useful measure of a model is increasingly how much capability fits within a given memory, compute and power budget, which makes local, private inference practical for work that would otherwise go to the cloud. The limits come with the approach: a compressed model cannot exceed the model it was made from, and the benchmark figures are PrismML's own. Hassibi has said the company intends to apply the technique next to models with several hundred billion parameters, where he expects compression to lose even less. It also sells custom work, tailoring Bonsai models to a customer's data and hardware.