SALOMI offers a research workspace exploring whether binary weight representations can match ternary baselines in transformer quantization. The project's key finding reveals that strict 1.00 bpp binary quantization fails as a strong solution for GPT-2 class language modeling under rigorous testing, providing realistic expectations for extreme quantization research.

More credible results cluster around 1.2-1.35 bpp using Hessian-guided VQ, mixed precision, or magnitude-recovery methods. The repository contains comprehensive quantization tools, runtime inference capabilities, evaluation kernels, and extensive documentation under Apache-2.0 licensing, prioritizing research exploration over polished production features.

The project embraces an honest assessment approach, preserving research chronology while providing curated reading paths through RESEARCH.md and key documents. This transparency offers valuable insights into both promising methods and where naive sub-1-bit claims break down, making it an important resource for researchers exploring practical limits of transformer quantization.