Building a Korean ambiguity solver fast enough to skip the GPU: 7,300 words/SEC

Cool product! (spent a year in Seoul as a student many moons ago, would've loved something like this)

If it's only 14M parameters, why not ship it to the client and run on the user's device?