Ternary neural network
AssessTechniques
A neural network that represents weights using -1, 0, and +1 values.
Why it's here
Placed in Assess: 2 article(s) of evidence from 2 source(s), led by research-stage coverage, with 2 in the last 30 days. Confidence 37%.
Evidence (2)
- 7The New Stack·9/17/2026researchIntel Compresses Ternary LLM Weights Below 1.58 Bits with BITCOS
Intel researchers introduced BITCOS, a storage format that compresses ternary model weights by separating zero positions from sign bits, without changing the model’s weights or outputs. In tests across multiple checkpoints and hardware platforms, it reduced storage to as low as 1.485 bits per weight and improved decoding throughput by up to 27%.
- 6Hacker News·9/16/2026researchResearch Breaks 1.58-Bit Barrier for Ternary LLMs
This arXiv paper reports a method for ternary large language models that exceeds the long-cited 1.58-bit per-weight barrier. The work is presented as a research advance in low-bit model representation and compression, with potential implications for more efficient inference and deployment.