Bonsai 2 27B: Near-Lossless Compression in a 9x Smaller Footprint

Bonsai 2 27B, an advanced AI model, has been released with a near-lossless compression and a significantly smaller footprint. This model, based on Qwen3.8 27B, offers enhanced reasoning, coding, vision, and agentic capabilities while maintaining a compact size ideal for local device deployment. Using ternary weights with FP16 group-wise scaling, the model achieves 1.76 effective bits per weight and a total footprint of 5.9GB.

Despite being 9 times smaller, Ternary Bonsai 2 27B retains 98.2% of the aggregate performance of its full-precision counterpart. This compression breakthrough enables the model to run efficiently on various devices, including NVIDIA GeForce RTX 5090 and M5 Max, consuming less energy and providing faster processing speeds. The release of Ternary Bonsai 2 27B marks a significant milestone in AI deployment, offering high-capacity models with a reduced memory footprint.

This development opens up possibilities for more efficient AI systems across different platforms, from personal devices to large-scale datacenters.