PrismML shrinks Bonsai 27B to a 3.9GB 1-bit build

PrismML released Bonsai 27B in a 3.9GB 1-bit build and a 5.9GB ternary build, both under Apache 2.0, per Latent Space's July briefing.

A 27-billion-parameter model that fits under 4GB runs on a laptop GPU or a decent phone, no cloud bill attached. That's the pitch: capability without the API meter running.

Ternary and 1-bit are both extreme quantization, the process of shrinking a model's weights down to fewer possible values per parameter. Ternary allows three states per weight, 1-bit allows two. Compressing that hard usually costs a model its sharper behaviors first, tool use, multi-step reasoning, image understanding. PrismML claims Bonsai keeps all of it.

Latent Space doesn't publish benchmark numbers alongside the claim, and Apache 2.0 licensing means anyone can run the model commercially without asking. The outlet grouped the release with the Hy3 launch as one trend: smaller models built to leave the cloud, not just fit inside it.

My take: 'claimed' is doing the heavy lifting here. A 27B model surviving 1-bit quantization with full agentic behavior intact would be a real result. But nobody's published eval numbers yet. Treat this as a size story until someone runs Bonsai through an actual tool-calling benchmark and posts the score.