> ## Documentation Index
> Fetch the complete documentation index at: https://docs.prismml.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Introduction

> Bonsai is a family of 1-bit and ternary language models from PrismML.

**Bonsai** is a family of open-weight models from [PrismML](https://prismml.com). 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 **3.8GB** versus 54GB for a full-precision model and the 1-bit Bonsai 8B model fits in **1.16 GB** versus **16.38 GB** 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.

<CardGroup cols={2}>
  <Card title="Bonsai (1-bit)" icon="bolt">
    Weights packed to a single bit each `{−1, +1}`. The smallest and fastest family.
  </Card>

  <Card title="Ternary-Bonsai (1.58-bit)" icon="scale-balanced">
    Weights take one of three values `{−1, 0, +1}`. Roughly 50% more the size of 1-bit and measurably higher quality.
  </Card>
</CardGroup>

## Why low-bit models

* **They run where full-precision models can't.** 1-bit Bonsai 27B needs only 3.8GB. 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 to 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.

<Warning>
  The speed and memory advantages depend on the runtime having **native low-bit kernels**. A runtime without them may dequantize the weights and silently lose the benefit. See [Formats & runtime support](/download/formats) before picking a tool.
</Warning>

## The lineup

Four sizes, each in both families, each under Apache-2.0:

| Model           | Parameters | Context        | 1-bit on disk (GGUF) | Ternary on disk (GGUF) |
| --------------- | ---------- | -------------- | -------------------- | ---------------------- |
| **Bonsai 27B**  | 27B        | 262,144 tokens | 3.9 GB               | 7.2 GB                 |
| **Bonsai 8B**   | 8.2B       | 65,536 tokens  | 1.16 GB              | 2.18 GB                |
| **Bonsai 4B**   | \~4B       | 32,768 tokens  | 0.57 GB              | 1.07 GB                |
| **Bonsai 1.7B** | \~1.7B     | 32,768 tokens  | 0.25 GB              | 0.46 GB                |

Bonsai 27B is the flagship and adds multimodal vision input. See the [model overview](/models/overview) for how to choose.

## Where to go next

<CardGroup cols={2}>
  <Card title="Quickstart" icon="play" href="/get-started/quickstart">
    Clone the demo repo and chat with Bonsai locally in one command.
  </Card>

  <Card title="Try it without installing" icon="browser" href="/get-started/try-online">
    Hosted GPU demo, in-browser WebGPU, Colab, and iOS apps.
  </Card>

  <Card title="Pick a model" icon="layer-group" href="/models/overview">
    Sizes, families, benchmarks, and what fits your hardware.
  </Card>

  <Card title="How the quantization works" icon="microchip" href="/models/quantization">
    What 1-bit and 1.58-bit actually mean, layer by layer.
  </Card>

  <Card title="Run locally" icon="terminal" href="/run/server">
    llama.cpp, MLX, the OpenAI-compatible server, and Open WebUI.
  </Card>

  <Card title="Integrations" icon="plug">
    Connect [Cline](/integrations/cline), [LM Studio](/integrations/lmstudio), [Ollama](/integrations/ollama), or [Hermes](/integrations/hermes).
  </Card>
</CardGroup>