PrismML's Bonsai 27B Brings Advanced AI Natively to Laptops and iPhones

PrismML has achieved a significant breakthrough in making advanced AI universally accessible, releasing Bonsai 27B, the first 27-billion-parameter-class model capable of running natively on everyday devices like laptops and even iPhones. This release is not a new pre-trained model but a highly optimized, low-bit representation of Alibaba's formidable Qwen3.6-27B, an open-weight multimodal language model known for its agentic coding and reasoning capabilities and a 262,144-token context window.

Bonsai 27B ships in two variants under the permissive Apache 2.0 license: a 1-bit binary version with a 3.9 GB footprint, designed to fit within the memory constraints of a mobile phone, and a ternary version utilizing {−1, 0, +1} weights at an effective 1.71 bits per weight, occupying 5.9 GB and optimized for laptop-class quality. This drastic reduction in size—down from approximately 54GB for a 16-bit Qwen3.6-27B or 18GB for a 4-bit version—transforms the landscape of on-device AI.

The implications are profound. By enabling sophisticated multi-step reasoning, structured tool calls, vision tasks, and coherent agentic loops to run locally, Bonsai 27B ushers in an era of enhanced privacy, reduced latency, and offline functionality. Data remains on the device, mitigating transmission risks, while eliminating cloud inference dependencies slashes operational costs and boosts real-time performance. This release challenges the long-held assumption that powerful AI necessitates vast cloud infrastructure, instead championing an "Intelligence Density" philosophy. It paves the way for a new wave of applications, making advanced AI a core, private, and efficient feature of consumer technology rather than a distant cloud service.