PrismML's Bonsai 27B: Qwen3-27B Runs on a Smartphone for the First Time

PrismML, a startup founded by researchers from the California Institute of Technology, has unveiled Bonsai 27B, which it calls the world's first 27-billion-parameter language model to run successfully on a smartphone. The technology relies on 1-bit and ternary weights, which compress Qwen3.6-27B by 9.4–14.2x while retaining 90–95% of the original quality.

How Bonsai 27B works

Bonsai 27B is not a new architecture. It is an extremely optimized version of Alibaba's open Qwen3.6-27B model. PrismML released two variants with different levels of quantization: a ternary one (weights take the values –1, 0 or +1) and a binary one (only –1 and +1). The method is called post-training quantization: the original weights are compressed after training, without retraining from scratch.

The specifics of the two versions:

• Ternary version: 5.9 GB, about 26 tokens per second on a standard laptop
• 1-bit (binary) version: 3.9 GB, fits in smartphone RAM
• Weight compression ratio: 9.4–14.2x relative to the original
• Quality retention: 90–95% of the full Qwen3.6-27B
• Developer: PrismML (founded by Caltech researchers)

The original Qwen3.6-27B in fp16 format weighs over 50 GB and requires a server-grade GPU. The 1-bit variant of Bonsai 27B is 13x smaller.

Why 1-bit weights change the rules

In standard language models, each parameter is stored in 16 or 32 bits. Moving to a 1-bit representation means a "weight" is literally a zero or a one. This sharply cuts memory use and speeds up inference on devices without a specialized GPU.

According to PrismML's publication, this approach is what allowed the model to run on an ordinary smartphone without a cloud connection. A year ago, that was considered unrealistic for models with 27 billion parameters. Until now, mobile LLMs were limited to models of 1–7 billion parameters, such as Google's Gemma 2B or Microsoft's Phi-3.5-mini.

"Bonsai 27B shows that the line between cloud and edge computing is rapidly disappearing,"

PrismML says in its official announcement.

Both model variants are openly available on Hugging Face, and developers can start experimenting now.

What this means

Running a 27-billion-parameter model on a smartphone is a significant technical threshold: powerful language models could run on billions of existing devices without cloud infrastructure, with full data privacy and in offline mode. With this result, PrismML has shown that 1-bit quantization has moved from an academic idea to a practical tool.

Frequently asked questions

Can Bonsai 27B run on an iPhone or Android?

The 3.9 GB 1-bit version runs on an ordinary smartphone without a cloud connection, which is exactly what PrismML demonstrated in its announcement.

How does Bonsai 27B differ from Qwen3.6-27B?

Bonsai 27B is Qwen3.6-27B with extremely compressed weights. PrismML applied 1-bit and ternary quantization, reducing the model's size by 9.4–14.2x while retaining 90–95% of the original's quality.