# Ternary Bonsai 2

> Family of 2B, 4B and 9B ternary language models for local use

- Category: Agents & dev (`agents`)
- Tags: `open-source`
- Official: https://prismml.com/news/bonsai-2-family
- Wunderlust: https://wunderlust.xyz/r/ternary-bonsai-2
- Image: https://wunderlust.xyz/media/ternary-bonsai-2-og.jpg
- Video preview: https://wunderlust.xyz/media/ternary-bonsai-2.webm
- Added: 2026-11-05

## What it is

Ternary Bonsai 2, by PrismML, are three compact models based on Qwen3.5 that store each weight as -1, 0 or +1. They weigh 0.43 GB (2B), 0.95 GB (4B) and 2 GB (9B).

## What it’s for

To run reasoning, vision and tool calling on PCs, phones and wearables, without depending on the cloud.

## Benefits

- Small sizes for local execution
- Support for NVIDIA via llama.cpp and Apple via MLX
- Weights under the Apache 2.0 license

## Caveats

- According to the maker, they keep about 92 to 100% of the original model's performance at 9B and reach up to 5x the throughput
- Benchmarks and speeds are reported by PrismML itself
- Likely match: associated by name and description

## Similar on Wunderlust

- [Ollama](https://wunderlust.xyz/r/ollama.md): Runtime for running language models on your own machine
- [Qwen3.8-Max](https://wunderlust.xyz/r/qwen3-8-max.md): Repository for Qwen3.8-Max, Alibaba Cloud's MoE model
- [Postiz](https://wunderlust.xyz/r/postiz.md): Open-source agentic scheduling tool for social media posts
- [Unsloth Studio](https://wunderlust.xyz/r/unsloth-studio.md): App for running, chatting with and training AI models locally
- [Paperclip](https://wunderlust.xyz/r/paperclip.md): Control center for agents organized like a company
- [MLX](https://wunderlust.xyz/r/mlx.md): Stack for running local models on Macs