A new arXiv preprint introduces MLFont, a machine-learning-driven approach to font subsetting and unicode-range slicing for web typography. The paper notes that CSS3 @font-face rules with unicode-range descriptors already partition web fonts into downloadable slices, but the efficiency of those slices varies widely by script.
Latin fonts are relatively light at around 20 KB in WOFF2 format, while Pan-European (LGC) and East Asian (CJK) fonts contain 2,297 to 30,759 codepoints, translating to file sizes from 158 KB up to 5.9 MB. That size gap makes low-latency loading particularly challenging for non-Latin fonts.
The proposed MLFont method applies machine learning to create range-compact slices, aiming to reduce the number of bytes a browser must download for a given text. The abstract is truncated, so full experimental details are not available, but the stated objective is clear: use learned models to improve on conventional slicing and subsetting, cutting latency for large font sets.