PrismML Launches Bonsai 2 27B, Its Most Capable Model Yet (prismml.com)

PrismML has launched Bonsai 2 27B, a groundbreaking 27-billion parameter model that significantly enhances reasoning, coding, vision, and agentic capabilities while being more compact and efficient. Leveraging the Qwen3.8 architecture, Bonsai 2 27B offers a substantial reduction in memory footprint, now at just 5.9 GB, which is over 9 times smaller than its full-precision counterpart. Despite this reduction, it retains over 98% of the performance of the previous model, showcasing rapid advancements in low-bit model efficiencies. This development allows sophisticated AI to be deployed locally, eliminating the traditional reliance on cloud infrastructure. The significance of this model lies in its ability to perform complex tasks across various modalities—encompassing reasoning, coding, and visual inputs—at an unprecedented scale and efficiency. Achieving an aggregate benchmark score of 83.9, the Bonsai 2 27B is designed for demanding applications such as local assistants and long-horizon tasks, emphasizing PrismML’s vision of “intelligence density.” By making powerful AI models more accessible and efficient, this launch has the potential to broaden the scope of AI application, fundamentally transforming how and where advanced models can be deployed. Users can download the model for free starting September 17, 2026, under the Apache 2.0 license.