Bonsai 27B: The First 27B-Class AI Model That Runs on a Smartphone

Short answer

Yes — a 27-billion-parameter AI model can now run on a smartphone. PrismML released Bonsai 27B on July 14, 2026, using extreme 1-bit and ternary quantization to squeeze Qwen3.6 27B into 3.9 GB without retraining.

How it works

Bonsai 27B uses ternary quantization (values -1, 0, +1) and 1-bit quantization to shrink the model from 54 GB to under 4 GB. This is a post-training compression technique that preserves most reasoning and vision capabilities.

Why it matters

Running a 27B model on-device means no cloud dependency, lower latency, and full privacy. Apple's Neural Engine on the iPhone 17 Pro provides 47 TOPS, enough to run the quantized model at usable speeds.

References

• PrismML - Bonsai 27B announcement