PrismML releases Bonsai 2 27B for local AI execution

Startup PrismML has released Ternary Bonsai 2 27B, a highly compressed large language model designed to run locally on personal computers and high-end mobile devices. Based on the Qwen3.8 27B model, the new version utilizes a “ternary weights” technique (+1, -1, or 0) combined with FP16 group-wise scaling. This method reduces the model footprint to just 5.9 GB, making it more than nine times smaller than its full-precision counterpart.

Despite the significant reduction in size, the model retains 98.2% of the original aggregate benchmark performance. The Bonsai 2 27B supports a 262K-token context window and multimodal text-and-image inputs. Led by Caltech academic Babak Hassibi, the technology aims to democratize AI access by eliminating cloud dependency, thereby enhancing user privacy and reducing operational costs. The model is released under the Apache 2.0 license.