Bonsai 2 27B: 5.9 GB of weights for a 27B model (twitter.com)

The release of Ternary Bonsai 2, a 27 billion parameter model weighing in at just 5.9 GB, represents a significant advancement in model efficiency for the AI/ML community. This compact model was evaluated against higher-memory counterparts, such as the full-precision Qwen 3.8 (54 GB) and Gemma 4 (12B QAT), showcasing its strong performance despite the drastic reduction in weight. Notably, Bonsai 2 scored 95% of the performance of the Qwen model on the IMO 2026 challenge, completing problems in approximately 70% of the time while using 22% more thinking tokens, signaling its potential for efficient reasoning in compact formats. This development highlights a crucial trend in AI towards enhancing computational efficiency while maintaining or improving reasoning capabilities. The ability to achieve high performance with significantly reduced memory footprint opens up new possibilities for deploying advanced AI models on resource-constrained devices or in applications where latency and power efficiency are critical. As Bonsai 2 demonstrates, achieving lower-weight models while retaining reasoning capacity can lead to practical applications in various domains, making this announcement a key moment for future AI advancements.