Developer ToolsJul 16, 2026
Bonsai 27B enables mobile LLM agents
PrismML released Bonsai 27B, a 27B-parameter LLM compressed to 1-bit and ternary weights, enabling local inference on mobile devices and browsers. This marks a significant shift towards persistent on-device agents.
Why now
This development is timely as it addresses the growing demand for efficient, local AI solutions, particularly for mobile and edge computing.
Key signals
Bonsai 27B demonstrates that 27B-class models can run on mobile devices, unlocking new applications for local agents.
Ternary quantization allows Bonsai 27B to achieve near-fp16 performance on limited memory, making it suitable for resource-constrained environments.
Bonsai 27B is running locally in browsers via custom WebGPU kernels, reducing the size from 54GB to 3.8GB.