The 1.58-bit shock

Ternary Bonsai has been released. Using {-1, 0, +1} ternary weights, it compresses down to 1.58-bit level, and it claims to reduce memory usage by about 9x compared to typical 16-bit models.

The model is provided in three sizes: 8B, 4B, and 1.7B. The announcement describes it as an extension of the recent 1-bit Bonsai, taking the approach of using slightly more capacity while minimizing performance degradation.

• Release form: Blog announcement and Hugging Face model collection
• Format support: Currently, MLX 2-bit packed format is the only packing format available
• Compatibility: 8B FP16 safetensors is also provided, allowing it to run with stock Hugging Face tools
• Performance claims: Introduced as outperforming most competing models of the same parameter class on standard benchmarks

This can be seen as an extension of the trend toward making ultra-low-bit LLMs practically usable in memory-constrained environments.