PrismML launches Bonsai 2 27B lightweight AI model

PrismML, a startup founded by Caltech researchers, has announced the release of Bonsai 2 27B, a second-generation ultra-lightweight multimodal generative AI model. The model utilizes a ternary weight approach—using only +1, 0, and -1 values—to compress Alibaba’s Qwen3.8 27B model from approximately 56 GB down to 5.9 GB.

Despite the significant reduction in memory footprint, Bonsai 2 27B retains high performance, achieving roughly 98% of the original model's capabilities. In various benchmarks, it remains highly competitive in coding, reasoning, and tool-calling tasks. The model is designed to run efficiently on consumer-grade hardware, such as Nvidia GeForce GPUs and Apple M-series chips, offering lower latency and improved data privacy by enabling local processing without cloud reliance.

PrismML has secured $22.25 million in seed funding from investors including Khosla Ventures, Cerberus Capital, and Caltech. The company is currently working on scaling this compression technology to models with hundreds of billions of parameters.