PrismML

PrismML is a U.S.-based AI company, founded by a group of Caltech researchers and led by Babak Hassibi, a Caltech professor.^{2, 4, 5, 6, 7, 8}

It presented itself publicly on 31 March 2026 as an AI lab whose work comes out of research developed at Caltech.^1, 2

Its first models were the 1-bit Bonsai 8B, 4B and 1.7B language models, released that day with open weights under the Apache 2.0 license.^{1, 2, 4, 9}

On 16 April 2026 it announced the Ternary Bonsai family in 8B, 4B and 1.7B sizes, which stores each weight on 1.58 bits using three values.³

On 26 May 2026 it released Bonsai Image 4B, a family of compact text-to-image models offered in 1-bit and ternary variants.^9, 10

On 14 July 2026 it announced Bonsai 27B, based on Qwen3.6 27B and offered in 1-bit and ternary variants.^11, 12

Bonsai 2 27B followed on 17 September 2026, a ternary model based on Qwen3.8 27B.^5, 8

On 15 October 2026 the company added Bonsai 2 models in 2B, 4B and 9B sizes, based on Qwen3.5, bringing the Bonsai 2 family to four sizes.⁴

The Bonsai 2 family replaces the first Bonsai generation, whose weights remain available on the company's Hugging Face organization.⁹

As of October 2026, Babak Hassibi is the founder and CEO of PrismML.^4, 7