There are some big speed claims for Rust-based TypeScript checkers. How do they compare with tsc on real codebases, and are they ready to use in CI?
How do the Rust TypeScript checkers compare with tsc on real projects?
Answers
One benchmark tested two projects on a Linux VM, using the median of five warm runs. The backend had roughly 1,120 files and the frontend roughly 1,160, mostly TSX.
tsc 7.0.2 took 21.0 seconds on the backend and 6.5 seconds on the frontend. Bun check 1.4.3 took 1.7 and 2.4 seconds, tsc-rs 0.7.0 took 1.4 and 2.4, tsrs 0.11.0 took 0.8 and 1.3, and tsc-ts 0.0.5 took 2.0 and 2.7.
Those are useful numbers, but they describe warm runs on two particular projects, not a general speed ranking for every codebase.
Reply to this answer
Speed and correctness need separate tests. All four caught the same three deliberately introduced errors as tsc in that benchmark, with matching messages. That's encouraging, but three errors won't tell you much about tricky generic inference, declaration emit, or project references.
For a trial, use a faster checker locally or on pull requests and keep tsc as a reference check in CI. Running it nightly would at least limit how long a missed error can go unnoticed.
Reply to this answer
For CI, check the release status and exit codes before chasing the fastest result. In the benchmark, tsrs was a development snapshot and returned exit code 2 on errors rather than 1, which could matter to scripts that expect a specific code.
Bun check was slower than tsrs there, but still cut the backend check from 21 seconds to 1.7. If you're already using Bun, that may be a more practical tradeoff. The source also described tsc-rs as experimental and tsc-ts as an early release.
Reply to this answer
I'd also compare against a newer tsc prerelease if you're evaluating these today. The baseline here was 7.0.2, while tsrs was built against 7.1 development code, so the comparison doesn't isolate the implementation from the version difference.
Reply to this answer
Add your answer
Answers are reviewed before publication.
n.396+tramnotch:
The benchmark didn't report memory usage, so the timings don't tell us how the tools compare on that front.