Prism ML Inc. has released Bonsai 2 27B, a multimodal generative AI model designed for local execution on consumer hardware. The model is based on Qwen3.8 27B and utilizes a ternary compression technique to reduce its size from approximately 56 gigabytes to 5.9 gigabytes.
The ternary method simplifies model weights into three values: +1, 0, and -1. According to the company, this approach allows the model to retain 98.2% of the original capabilities of Qwen3.8, which typically requires 9.4 gigabytes of memory at its minimal footprint.
Performance benchmarks show Bonsai 2 scoring 81.6 in coding tasks and 82.7 in knowledge and reasoning. The model achieves 143 tokens per second on an Nvidia GeForce GTX 5090 and 46.8 tokens per second on an Apple M5 Max chip, consuming 0.714 megawatt-hours per token.
The model is compatible with Nvidia GPUs via CUDA and Apple devices, including Mac, iPhone, and iPad, through MLX. Prism ML has made the model weights available under the Apache 2.0 license to enable local AI processing without third-party data sharing.
