PrismML's tiny LLM Bonsai 2 matches 98% of Qwen's benchmark scores
prismml releases bonsai 2, a 5.9 gb llm that matches 98% of qwen's benchmark scores, enabling ai on pcs and smartphones. this could change how we use ai.
prismml, founded by caltech researchers, has developed a compression technology that enables large language models to run on pcs and smartphones. its latest model, bonsai 2 27b, compresses qwen3.8 27b down to 5.9 gb while matching 98% of its benchmark scores. this is a 9x to 10x reduction in memory compared to the original model. prismml's approach, called ternary weights, simplifies model weights from 16 bits to three: +1, −1, or 0. the startup's goal is to apply this technology to even bigger models, with plans to release models in the several-hundred-billion-parameter range in the next couple of months.