Downright 1.0.18 spends less time parsing, diffing, and updating your document. These benchmarks compare the same workloads on the same Mac.
Measured improvements
| Measurement | Before | After | Less time |
|---|---|---|---|
| cmark conversion | 13.076 ms | 10.605 ms | 18.9% |
| Full parse | 27.477 ms | 20.807 ms | 24.3% |
| AST diff | 0.030 ms | 0.018 ms | 40.0% |
| Text diff | 1.605 ms | 0.503 ms | 68.7% |
| Incremental decoration | 0.104 ms | 0.102 ms | 1.9% |
| Wholesale decoration | 157.616 ms | 73.180 ms | 53.6% |
| Typing/storage | 0.149 ms | 0.099 ms | 33.6% |
| Semantic convergence | 28.090 ms | 20.483 ms | 27.1% |
| Parse 100 KB | 23.020 ms | 16.489 ms | 28.4% |
| Syntax highlighting | 0.090 ms | 0.065 ms | 27.8% |
Apple M4 / macOS 27.0. Test conditions and source.
Test conditions and limits
Apple M4 / macOS 27.0, measured 2026-10-27. p95 latency, fastest baseline versus slowest result across three interleaved pairs on the same Mac. The corpus is 120,825 characters / 4,885 lines; the separate 100 KB parse uses exactly 100,000 characters.
The measured code is included in Downright 1.0.18 (build 356). Raw runs and source identity are available alongside the performance record.
These timings cover parsing, diffing, storage, and decoration. They do not measure first-pixel launch time, visible scrolling, or complete window transitions.
Parsing, search, and memory
| Measurement | Before | After | Less time / memory |
|---|---|---|---|
| Unicode inline-heavy parsep95 | 48.811 ms | 11.212 ms | 77% |
| Fuzzy heading searchSingle invocation | 11.036 ms | 0.912 ms | 92% |
| Peak physical footprintTwo fresh-process runs | 449.2 MiB | 204.1–252.2 MiB | 44–55% |
Apple M4, macOS 27.0 (26A5406e). Test conditions and source.
- Unicode inline-heavy parse: 16,800-byte Unicode line with repeated formatting and links. p95.
- Fuzzy heading search: 12-character query against a 1,500-character heading. Single invocation.
- Peak physical footprint: 1,032,709-byte document; edit, save, undo, scroll, and presentation switches. Two fresh-process runs.
The Unicode result is a p95 latency. The search result is a single measured invocation, so it does not establish a latency percentile. The memory result is an observed range across two fresh processes, measured with the kernel's lifetime maximum physical-footprint counter.
The below-150-MB interactive-session target remains unverified. The two-run range is not a percentile or a guaranteed ceiling. A later 126.2 MB observation measured opening the stress document only; it does not replace the interactive-session measurements above. Physical footprint and live heap allocations measure different things.
The memory workload uses a document with 57,043 lines and 16,298 tasks. Each run opens a fresh isolated native host, enters Source, inserts seven characters, saves, undoes and saves again, scrolls down and up, and performs three further Source/Document cycles. The document's original bytes were restored exactly.
Historical pipeline baseline
- Corpus: 120,825 characters / 4,885 lines
- Date: 2026-08-06
- Qualification: Historical baseline only. Not a cross-machine claim.
- Not measured: TextKit layout, scroll frames, IME input, live window frame time
The historical small-document parse used roughly 48 KB, despite its original 100 KB label. It is not a 100 KB measurement or a first-pixel timing. The existing timings are preserved; the workload correction prevents comparing that row with a genuine 100 KB test.
| Measurement | p50 | p95 | Target |
|---|---|---|---|
| cmark parse | 12.827 ms | 13.309 ms | Informational |
| Full MarkdownParser.parse | 29.463 ms | 30.372 ms | Informational |
| AST dirty set, one-character edit | 0.030 ms | 0.031 ms | Informational |
| Text diff, external rewrite | 2.099 ms | 2.164 ms | Informational |
| Incremental decoration | 0.097 ms | 0.104 ms | <8 ms |
| Wholesale decoration | 103.722 ms | 171.493 ms | Informational |
| Source edit and paragraph map | 0.146 ms | 0.153 ms | <8 ms |
| End-to-end semantic convergence | 30.583 ms | 31.445 ms | <100 ms |
| Small-document parse (~48 KB) | 12.078 ms | 12.395 ms | Informational |
| Syntax highlight, 10 KB Swift | 0.093 ms | 0.094 ms | Informational |
How to use the numbers
Use the benchmark to detect regressions in the same corpus and build path. Do not treat it as a promise about every Mac, document, display refresh rate, or window size. The app's release checks keep the methodology and limitations visible.