Bonsai 27B (1-bit LLM): The First 27B-Class Model to Run on a Phone (prismml.com)

Bonsai has announced the launch of Bonsai 27B, a groundbreaking 27 billion-parameter model that is the first of its kind to run directly on a mobile phone. Building on earlier advancements in low-bit weight models, Bonsai 27B incorporates two variants: Ternary Bonsai 27B, which uses ternary weights and operates efficiently on laptops, and 1-bit Bonsai 27B, designed specifically for the memory constraints of smartphones like the iPhone 17 Pro. The 1-bit variant reduces the model's size to just 3.9 GB, enabling complex capabilities such as multi-step reasoning and coherent tool calls on portable devices. This ensures that advanced AI functionalities are no longer confined to cloud processing, addressing privacy concerns while maintaining high performance. The significance of Bonsai 27B extends beyond mere size. As AI workloads increasingly demand sustained interactions—a shift from single responses to ongoing processes—having a powerful, locally running model allows for more sophisticated, agent-like behavior without the need to rely on cloud-based infrastructure. This paradigm shift opens new avenues for deploying AI effectively while safeguarding user data. With Bonsai 27B's enhanced intelligence density, developers can leverage advanced on-device capabilities that redefine the landscape of personal AI assistants and offline applications. The model is set to influence the economics of AI deployment, promising a future where robust AI solutions are accessible on everyday devices.