PrismML released Bonsai 2 27B on September 17, compressing Qwen3.8 27B into a package the company says occupies 5.9 GB. The goal is to make capable reasoning models practical on devices with less memory.
PrismML reports that Bonsai 2 retains 98% of the original model’s aggregate benchmark score. That percentage compares the published test results; it does not mean the compressed model answers 98% of all questions correctly, or that no individual task loses quality.
A smaller package can make a checkpoint easier to store and load, but inference also needs working memory for the runtime and conversation. The reported file size is not a verified minimum specification for a phone or PC.