A 41,000-parameter model beats Shiki at code highlighting

Shu Ding at Vercel Labs published gpu-lexer on September 9, a syntax highlighter that skips per-language grammar files entirely. Instead it splits source code into words, whitespace, newlines, and symbols, then labels each piece with a model trained on 4,688,781 tokens. The whole thing runs on WebGPU and fits in 41,321 parameters, 27.4KB, and covers 75 languages.

The benchmark is where it gets interesting. On a stress test built from three.min.js concatenated ten times over, gpu-lexer finished a warmed browser run in 402ms. Shiki took 29.6 seconds on the same input. Highlight.js took 1.29 seconds. That's a 73x gap against the tool most sites already use.

Speed isn't free here. gpu-lexer agrees with Shiki's token labels 88.02% of the time on held-out files, and 90.35% weighted across popular languages. Shiki builds on TextMate grammars, hand-written rules tuned over years for exact tokenization. A model trained on token statistics doesn't reproduce that exactness. It approximates it, fast.

For most code blocks on a blog or docs site, a highlighter that's occasionally wrong about a token boundary and imperceptibly different to a human reader is a fine trade for near-instant rendering, especially on large files or repeated highlighting passes where Shiki's grammar-walking cost adds up. For anything that needs exact tokenization, like a linter or an editor's semantic highlighting, the accuracy gap is the whole ballgame and a trained model isn't a drop-in replacement.

What gpu-lexer doesn't say is how it holds up on languages outside the tested set, or what happens on adversarial input designed to break the labeling model rather than a hand-written grammar.