The 'Ternary Bonsai 2 27B' is an AI model that reduces the size of the Qwen3.8 27B to 5.9GB while maintaining 98.2% of its performance.

PrismML, an American company, has announced the ' Ternary Bonsai 2 27B, ' a model based on Alibaba's 'Qwen3.8 27B' that achieves high local throughput and superior energy efficiency.

PrismML — Introducing Bonsai 2 27B: Near-Lossless Compression in a 9x Smaller Footprint

Today, we're announcing Ternary Bonsai 2 27B.

Based on Qwen3.8 27B, Bonsai 2 27B is 9x smaller than its full-precision counterpart while retaining 98.2% of its aggregate benchmark performance.

Two months after the first Bonsai 27B release, the biggest change is quality. The… pic.twitter.com/R8koycSkmJ

— PrismML (@PrismML) September 17, 2026

The Ternary Bonsai 2 27B is a model that combines powerful inference, coding, and agent capabilities with the significantly smaller memory footprint, high local throughput, and excellent energy efficiency characteristic of the 'Bonsai' series.

PrismML's previously released 'Bonsai 27B' boasted significantly reduced memory usage, making it compatible with iPhones. This new Bonsai 2 27B also features compatibility with Apple devices (Mac, iPhone, iPad), including iPhones.

'Bonsai 27B,' an AI model with 27 billion parameters that can be run locally on an iPhone, has been released. It's a smaller version of Qwen3.6-27B, using only 3.9GB of memory, allowing it to run at a practical speed on a smartphone alone - GIGAZINE

It's based on the Qwen3.8 27B, but its size is 5.9GB, which is one-ninth the size of the Qwen3.8 27B. Despite being a large 27B model, it's small enough to be considered for local execution even on graphics cards with only 8GB of VRAM. Furthermore, it achieves 98.2% of the benchmark scores shown by the Qwen3.8 27B, demonstrating a significant advantage of 'maintaining performance while reducing size.'

Compared to the previous generation Bonsai 27B, the Bonsai 2 27B achieves the following:
- The base model has been upgraded from Qwen3.6-27B to Qwen3.8 27B.
- High performance retention rate of 98.2% compared to the full-precision model.
- Improved inference, coding, vision, and long-term agent performance

A comparison table of various benchmarks is shown below.

Ternary Bonsai 2 27B is released under the Apache 2.0 license. It is also available on Hugging Face.

Bonsai-2 - a prism-ml Collection