PrismML Unveils Compact LLM for On-Device AI

PrismML, an AI lab, has introduced Bonsai 2 27B, its latest compact large language model. This model significantly compresses Qwen3.8 27B, an open-source model from Alibaba, to enable its operation on personal computers and high-end smartphones.

The Bonsai 2 27B model occupies 5.9 GB of memory, representing a nine to ten-fold reduction compared to the original Qwen3.8 27B. It achieves 98% of Qwen's aggregate benchmark scores. The company previously raised a $22.25 million seed funding round. Its initial Bonsai model garnered over 11 million downloads, with other smaller models accumulating an additional 2.6 million downloads.

PrismML was established by a group of researchers from Caltech and is led by CEO Babak Hassibi, a Caltech professor recognised for his expertise in compression technologies. Ion Stoica, a co-founder of Databricks and director of Berkeley's Sky Computing Lab, serves as an advisor. The startup has secured investment from Khosla Ventures, Cerberus Capital, and Caltech.

PrismML's technology aims to democratise access to powerful reasoning models by making them small enough for on-device use, ensuring privacy and potentially free access for users. The company asserts that its compression method maintains virtually identical performance to the original, uncompressed models.

The company intends to apply its compression techniques to even larger models, potentially in the several-hundred-billion-parameter range, in the coming months. CEO Babak Hassibi suggests that larger models may offer greater scope for compression without compromising intelligence, potentially leading to 100% benchmark parity.