Bonsai 2 27B: Near-Lossless Compression in a 9x Smaller Footprint (prismml.com)

PrismML has unveiled Ternary Bonsai 2 27B, a cutting-edge multimodal AI model that achieves near-lossless compression with a 9x smaller memory footprint compared to its full-precision counterpart. Building on the earlier Bonsai 27B model, this new iteration integrates ternary weights and FP16 group-wise scaling, resulting in a compact 5.9GB size while retaining an impressive 98.2% of aggregate performance across various benchmarks. With enhancements in reasoning, coding, vision, and agentic capabilities, Ternary Bonsai 2 27B is now more adept for real-world applications, particularly in domains requiring high computation efficiency and fast processing speeds. The significance of this release lies in its ability to democratize access to advanced AI technologies on local devices, enabling applications such as private document analysis, multimodal debugging, and sophisticated coding workflows without relying heavily on cloud resources. Its energy efficiency and high throughput—achieving up to 143 tokens/second on high-end NVIDIA GPUs—make it ideal for tasks requiring rapid iteration. By pushing the boundaries of low-bit technology, Ternary Bonsai 2 27B presents a promising direction for future AI deployment strategies, potentially reshaping the economics and architecture of AI systems across a range of devices and environments.