June 6, 2026 · 18 min read
Building a Korean ambiguity solver fast enough to skip the GPU: 7,300 words/sec
A 14M-param KoELECTRA quantized to int8, hand-rolled in pure Rust, resolves ~7,300 Korean lemma ambiguities/sec on a plain CPU, no GPU. The four attempts it took to get there.
lemmatizermachine-learningrustkoreaninferenceint8