Bonsai 27B от PrismML: модель Qwen3-27B впервые запустили на смартфоне

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

How Bonsai 27B Works

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

Specific parameters of the two versions:

• Ternary version: 5.9 GB, speed of approximately 26 tokens per second on a standard laptop
• 1-bit (binary) version: 3.9 GB, fits in a smartphone's 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 version of Bonsai 27B is 13 times 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 that a "weight" is literally a zero or a one. This radically reduces memory footprint and accelerates inference on devices without a dedicated GPU.

According to PrismML's publication, this approach is what made it possible to run the model on an ordinary smartphone without a cloud connection — something that was considered impossible just a year ago for models at the scale of 27 billion parameters. Previously, mobile LLMs were limited to models of 1–7 billion parameters: Google's Gemma 2B or Microsoft's Phi-3.5-mini.

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

PrismML announcement.

Both model variants have been published as open access on Hugging Face, and developers can start experimenting right now.

What This Means

Running a 27-billion-parameter model on a smartphone is an important technological milestone: powerful language models will be able to run on billions of existing devices without cloud infrastructure, with full data privacy and in offline mode. With this result, PrismML has confirmed 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 1-bit version weighing 3.9 GB runs on an ordinary smartphone without a cloud connection — this 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 size by 9.4–14.2x while retaining 90–95% of the original quality.