PrismML releases compressed Bonsai 2 27B model — TechCrunch
Artificial intelligence laboratory PrismML has released the Bonsai 2 27B language model, a compressed version of Alibaba’s open Qwen3.8 27B model. TechCrunch reports that the model is 5.9 GB in size. According to PrismML’s estimate, this makes it possible to run it on personal computers and possibly on high-end smartphones. The model requires 9–10 times less memory than the original Qwen3.8 27B.
Benchmark results
According to PrismML, Bonsai 2 27B achieves 98% of Qwen’s aggregate benchmark score. The first version of Bonsai, released in March, achieved 95% of this figure. The company said that the first Bonsai model has been downloaded more than 11 million times, while its smaller models have received another 2.6 million downloads.
PrismML was founded by a group of researchers from the California Institute of Technology. The company is led by Caltech professor Babak Hassibi, who specializes in compression technologies. The startup’s adviser is Databricks co-founder and director of the Sky Computing Lab at the University of California, Berkeley, Ion Stoica. PrismML’s investors include Khosla Ventures, Cerberus Capital, and Caltech.
More current news is available on the UA.News Telegram channel Telegram.
Ternary weights
PrismML’s technology is based on reducing the amount of data needed to store a model’s “weights,” the information it learns during training. Typically, 16 bits are used for each weight. Under the company’s approach, called ternary weights, values are reduced to three options: +1, −1, or 0. This reduces the amount of memory required for the model.
Hassibi said he hopes PrismML will release models with several hundred billion parameters over the next few months. According to him, larger models may offer more opportunities for compression while preserving their intelligent characteristics. Stoica noted that running powerful models directly on users’ devices could improve data privacy because the data would not need to be sent to the cloud.
