27.5KB WebGPU-based syntax highlighter guesses token types without language grammars

The highlighter, called gpu-lexer, splits source code into words, whitespace, and symbols, then uses a small WebGPU model to label each part. Instead of matching against a grammar, the model guesses the type of each token from its neighbours and the overall file. This lets it work on any language, even ones it was never trained on, but it may mislabel tokens that look unfamiliar.

Bundle size is the most visible change. Existing highlighters like Shiki or Starry Night ship hundreds of kilobytes of grammars; gpu-lexer ships one 27.5 KB model that covers every language. The trade-off is accuracy: on held-out files, 12.57 % of its labels differ from Shiki’s. The difference may grow for languages outside the training set or for code that uses unusual syntax.

Performance is another win. On a 5.56 M-character file, gpu-lexer took 471.7 ms, while Shiki took 30.4 s. The WebGPU model runs in parallel on the GPU, so it scales better with file size. However, the model still needs a warm-up run, and the first highlight may take longer. The benchmark excluded DOM rendering, so real-world latency will depend on how the spans are painted.

The model is not a drop-in replacement. It returns spans with nine coarse classes instead of fine-grained grammar tokens. Tools that rely on exact token types, like linters or refactoring engines, will need to map the classes back to their own schemas. The experiment also only measured agreement with Shiki, not objective correctness, so the 90 % score may not hold for all codebases.