Model releases & benchmarks
Breaking the 1.58-bit Barrier for Ternary LLMs
The captured HN discussion weighs a reported move from 1.58 to 1.48 bits per ternary weight against questions about competing quantization methods and whether the format works in memory.
Why it matters: If the reported packing works for the intended use, fewer bits per weight could matter for model storage and constrained devices. The discussion also flags a key practical question: whether the compact encoding can be used during inference or must first be expanded.