PrismML's Bonsai: First 27B AI Model That Fits on Smartphone

On July 15, 2026, PrismML introduced Bonsai 27B, which is notable for being the first 27-billion-parameter AI model that can operate on a smartphone. This model is compressed to only 3.9 GB, allowing it to run efficiently on an iPhone 17 Pro Max at a speed of 11 tokens per second. In AI terminology, a token represents the basic unit of information that models can process.

Typically, 27-billion-parameter models require around 54 GB of memory to function at half precision, making them impractical for most consumer devices. However, Bonsai achieves this capability by utilizing a novel compression technique that significantly reduces the model's memory requirements without sacrificing performance. The ternary variant of Bonsai, slightly larger at 5.9 GB, can achieve 26 tokens per second on an M5 Pro laptop, further demonstrating its efficiency.

This breakthrough has implications for AI deployment, as it might enable resource-constrained environments like smartphones to utilize advanced AI capabilities. The compression technology is rooted in Caltech intellectual property, targeting the reduction of each model's weight from 16 bits of standard precision to a more compact representation. This not only minimizes space but also allows the model to maintain about 94.6% of its full-precision benchmark performance.

According to reports, Apple is reportedly in early discussions with PrismML regarding the underlying compression technology for future potential applications, indicating a keen interest from major tech companies in this innovation.

Why It Matters

• The ability to run such powerful models on consumer devices broadens accessibility to advanced AI tools.
• PrismML's approach contrasts with traditional low-bit models, as they have effectively managed to compress the entire model end-to-end, preserving performance while using less memory.
• If successful in collaboration with Apple, PrismML's technology could significantly influence smartphone capabilities regarding AI.

What to Watch

• Monitor potential collaborations between PrismML and major tech companies like Apple for future AI applications.
• Keep an eye on the development of new models leveraging similar compression techniques to expand capabilities in mobile devices.
• Watch how competitors respond to Bonsai's capabilities in the AI model landscape.