June 4, 2026 Building a Korean ambiguity solver fast enough to skip the GPU: 7,300 words/sec
How Kimchi Reader's Korean ambiguity solver, a 14M-parameter KoELECTRA quantized to int8, ended up running server-side on a plain CPU with no GPU, resolving thousands of word-sense ambiguities per second. The four attempts it took to get there.