Bonsai 2: A 27B AI That Runs in Your Browser for Free

A new AI model called Ternary Bonsai 2 showed up on Hugging Face, a popular public library where people share AI models. It's built from a 27-billion-parameter model — think of those as 27 billion tiny dials the AI adjusts to learn things. Normally, a model that size takes up around 54GB of space and needs a serious computer or an expensive cloud account to run. This version squeezes into under 6GB, which is roughly the size of a couple of movies, and it can run right inside a web browser.

The trick is something called "ternary weights." Instead of storing each of those 27 billion dials as a long, precise decimal, the model only stores three possible values: minus one, zero, or one. It's like describing a photo using just three shades instead of millions — you lose some fine detail, but you can still tell what you're looking at. The model card claims this keeps 98.2% of the original model's quality, while making the file nine times smaller than the standard format.

Why does that matter to you? Because it means AI could live on your own device instead of someone else's server. Nothing you type gets sent to a company, there's no monthly bill, and it can work without internet. Browser makers have recently added a technology called WebGPU, which lets websites borrow your graphics chip — the part of your computer built for games — to do heavy AI math. A live demo is already online, so in theory you can try it today without installing anything.

The catch: that 98.2% figure is the creators' own claim, not an independent test. Browser-based AI also tends to be slower and hungrier for battery than a cloud service. And Bonsai 2 comes from a small community group, not Google or OpenAI, so it hasn't been battle-tested by millions of users. Expect a capable, private assistant for everyday questions — not a replacement for the biggest paid tools on hard tasks.