PrismML has unveiled its next generation of AI models, the Ternary Bonsai 2 series, designed to bring sophisticated artificial intelligence capabilities directly to mobile devices, computers, and even wearable electronics. This new series includes three versions: 2B, 4B, and 9B, aiming to enable local execution of tasks like image recognition, tool invocation, and general command processing without constant reliance on cloud services.
Revolutionary Architecture
The Ternary Bonsai 2 models are built upon Alibaba’s Qwen 3.5 architecture. A core innovation is their use of a ‘ternary’ representation for weights, meaning model parameters are encoded using only three values: -1, 0, and +1. This optimization significantly reduces model size. According to PrismML, the 2B, 4B, and 9B versions occupy approximately 0.43 GB, 0.95 GB, and 2 GB respectively, making them ideal for deployment on devices with limited memory.
Practical Value for Users
The Bonsai 2 series emphasizes the practical utility of local AI agents. These models are engineered to follow instructions precisely, handle everyday tasks, interact with mobile device tools, and avoid repetitive responses. PrismML tested the models across six categories: tool invocation, mathematical reasoning, instruction following, coding, knowledge and logic, and visual perception. Test results indicate that the 9B model requires about 40% fewer tokens to complete standard instructions and mathematical tasks compared to previous versions.
High Performance and Energy Efficiency
PrismML reports substantial performance gains. On powerful hardware like the RTX 4090 and M5 Pro, the new models achieve up to 5 times higher throughput and 3 times better energy efficiency compared to models using standard weight representations. The results on mobile devices are particularly impressive. For instance, on an iPhone 15 Pro Max, the 9B model can process up to 34 tokens per second, while the smaller 2B model reaches a remarkable 120 tokens per second. This opens up new avenues for developing powerful AI applications directly on smartphones.
Ternary Bonsai 2 represents a step towards truly autonomous artificial intelligence that operates where it’s needed, without network dependency.
A Complete Range of Solutions
Combined with the previously released Bonsai 2 27B model, the new series creates a comprehensive product line capable of handling a wide spectrum of tasks, from the simplest to the most complex workflows. This allows developers to select the optimal option for their specific projects.
Availability and Licensing
Developers can easily integrate the Ternary Bonsai 2 models into their projects. They are available for running on NVIDIA GPUs via llama.cpp, and can also be deployed on Mac, iPhone, and iPad using MLX. Notably, the model weights are released under the Apache 2.0 license, granting freedom for both commercial and non-commercial use and modification.
