Skip to main content
Bonsai is a family of open-weight models from PrismML. Instead of the 16-bit weights most models ship with, Bonsai models are quantized to 1-bit (weights in {−1, +1}) or ternary (weights in {−1, 0, +1} ). The practical consequence: the 1-bit Bonsai 27B model fits in 3.9GB versus 54GB for a full-precision model and the 1-bit Bonsai 8B model fits in 1.16 GB versus 16GB for the same architecture in FP16. That is a 14x reduction, which is what lets Bonsai run on hardware that could never load a full-precision model of the same class.

Bonsai (1-bit)

Weights packed to a single bit each {−1, +1}. The smallest and fastest family.

Ternary-Bonsai (1.58-bit)

Weights take one of three values {−1, 0, +1}. Roughly 50% more the size of 1-bit and measurably higher quality.

​
Why low-bit models

  • They run where full-precision models can’t. 1-bit Bonsai 27B needs only 3.9GB. It loads and runs on an iPhone 17 Pro, while leaving some memory for KV Cache. A 16-bit 27B model needs 54GB for the weights alone.
  • They generate faster. Token generation is memory-bandwidth-bound. Moving 1-1.58 bits per weight instead of 16 means each token requires a fraction of the memory traffic, so decode speed scales up accordingly on the same hardware.
  • They use less energy. Measured at roughly 4 to 5x better energy per token than 16-bit counterparts (0.068 mWh/tok for 1-bit 8B on an iPhone 17 Pro Max versus 0.132 mWh/tok for 4-bit 8B model).
  • They keep data local. Inference runs entirely on your machine. Nothing leaves it.

​
The lineup

Five models, each under Apache-2.0: Bonsai 27B is the flagship and adds multimodal vision input. See the model overview for how to choose.

​
Where to go next

Quickstart

Clone the demo repo and chat with Bonsai locally in one command.

Try it without installing

Hosted GPU demo, in-browser WebGPU, Colab, and iOS apps.

Pick a model

Sizes, families, benchmarks, and what fits your hardware.

How the quantization works

What 1-bit and 1.58-bit actually mean, layer by layer.

Run locally

llama.cpp, MLX, the OpenAI-compatible server, and Open WebUI.

Integrations

Connect Hermes, or OpenClaw.