Introducing Ternary Bonsai: Top Intelligence at 1.58 Bits
Summary
PrismML announces Ternary Bonsai, a 1.58-bit language model family designed for strong performance at a dramatically reduced memory footprint. Available in 8B, 4B, and 1.7B sizes, it uses ternary weights with a shared scale to achieve ~9x memory reduction versus 16-bit models while delivering competitive benchmarks and improved energy efficiency. The release emphasizes on-device compatibility on Apple devices, highlights throughput gains, and provides access to a whitepaper and the Apache 2.0-licensed weights.