Introduction
PrismML's announcement of the Bonsai 27B model marks a significant breakthrough in artificial intelligence. For the first time, a 27B-class model can run directly on a phone, notably an iPhone 17 Pro. This opens revolutionary possibilities for developers and businesses looking to integrate advanced AI capabilities into portable devices.
What Makes Bonsai 27B Unique
The Bonsai 27B, based on the Qwen3.6 27B model, is designed to be multimodal, meaning it can process both textual and visual data. This model comes in two variants:
- Ternary Bonsai 27B: Uses ternary weights with FP16 scaling, requiring only 5.9 GB. It is the quality-oriented version, ideal for laptops.
- 1-bit Bonsai 27B: Utilizes binary weights for a footprint of just 3.9 GB, suitable for smartphones.
Both variants retain reasoning, tool-calling, and agentic processing capabilities.
Performance and Efficiency
Despite compression, the Bonsai 27B maintains much of the intelligence of the full-precision model. Benchmarks show that the ternary model retains 95% of the baseline model's accuracy, while the 1-bit version retains 90%. Here are some key figures:
- Mathematics: 93.4% for the ternary model, 91.7% for the 1-bit model.
- Coding: 86.0% for the ternary model, 81.9% for the 1-bit model.
- Instruction Following: 71.8% for the ternary model, 65.8% for the 1-bit model.
Potential Applications
With its ability to run on a phone, the Bonsai 27B opens new opportunities across various domains:
- Mobile Applications: Integrate advanced features like image recognition or real-time text analysis.
- IoT and Wearables: Provide local processing capabilities without relying on the cloud.
- Accessibility: Enhance tools for disabled individuals through smarter, more responsive applications.
Conclusion
PrismML's Bonsai 27B is an innovation that could redefine the boundaries of mobile AI. By reducing model sizes while retaining much of their capabilities, PrismML paves the way for wider adoption of AI in portable devices.
Let's discuss your project in 15 minutes.