Half the Weights in Ternary LLMs Are Zeros, and Exploiting That Beats the 1.585-Bit Floor

arXiv 2609.16338·medium signalXBlueskyLinkedInCopy link

Breaking the 1.58-bit Barrier for Ternary LLMs (arXiv 2609.16338, submitted 14 Sep 2026) measured the actual symbol distribution across 29 ternary models and found zeros account for up to 51.5% of all weights, invalidating the equiprobable assumption behind the log2(3) ≈ 1.585-bit reference and the 1.625 bits/weight that five-trit packing rounds to in practice. Their BITCOS layout, a dense presence bitmap plus a compacted sign vector, costs 2 - z bits per weight at zero density z. It stores weights more compactly than five-trit packing in 26 of 29 tested models and reaches 1.485 bits per weight on the sparsest, with an unpacking scheme designed for modern processors.