Parser benchmark results

Local prototype measurements. Run timestamp: 2026-11-08T12:19:21.836Z.

Runtime: v22.22.0, win32/x64. Baseline: markdown-it 14.3.3. Chunk size: 64 UTF-16 code units.

Batch and streaming speed ratios

Workload	Source code units	markdown-it batch ms	XMarkdown batch ms	Batch speed ratio	Stream vs repeated XMarkdown	Stream vs repeated markdown-it
short prose	510	0.03	0.05	0.64×	1.51×	1.43×
long prose	27,751	1.03	1.26	0.82×	74.79×	64.20×
code and tables	12,900	0.56	0.67	0.85×	22.89×	22.59×
long list	16,000	1.30	1.36	0.96×	29.15×	28.10×
graphics report	17,750	0.45	0.56	0.81×	41.75×	Different features

Batch ratio is baseline time / XMarkdown time: less than 1 means XMarkdown is slower. Streaming ratios compare a complete workflow of reparsing every growing prefix with section previews and one final full parse. They are not single-parse speedups.

Complete streaming workflow timings

Workload	Repeated XMarkdown ms	Incremental XMarkdown ms	Repeated markdown-it ms	Chunks	Incremental submitted input
short prose	0.13	0.09	0.13	8	2.67N
long prose	306.62	4.10	263.18	434	2.99N
code and tables	62.59	2.73	61.77	202	2.87N
long list	152.14	5.22	146.69	250	3.40N
graphics report	71.98	1.72	—	278	2.50N

Method and limitations

Reused parser instances and bounded extension cache; hydrated tokens, no HTML rendering; 60 interleaved batch warmups, two streaming warmups, seven timed trials with rotating condition order; median per operation. Streaming compares full-prefix parsing on every chunk with provisional section patches plus one full final parse. Repeated graphic specifications can hit the cache in all XMarkdown conditions. Graphics baseline has unequal features; only the same-XMarkdown stream comparison is used there. See features-report.md for cold and cached rendering costs.

Both parser instances are reused. Batch warmups and trial order are interleaved to reduce order bias. Times are medians of seven trials, with minimum and maximum values retained in results.json.
Functions parse inline content into tokens. The benchmark excludes HTML generation, SVG layout, DOM updates, networking, model generation, and AI token cost.
The plain-workload Markdown core uses the same options. XMarkdown additionally applies its extension rules, validation policy, and diagnostics. The graphics workload has extra XMarkdown capabilities, so its bare markdown-it time is contextual rather than equivalent-feature performance.
Repeated-prefix parsing produces a whole document tree after every chunk. Incremental parsing emits fewer provisional section states and returns a complete final tree. These are deliberately different update strategies, with different preview timing.
Paragraph previews may lag by a block; long ambiguous containers can defer. The final streamed output is checked against batch output by the test suite.
Corpora are synthetic and repetitive, from 510 to roughly 28,000 code units. Results are a local pilot, not evidence about every Markdown parser, document, device, model, or chunk size.
A 50× streaming result on a workload does not establish a 50× batch parser or a 50× faster AI response. Timing variation and chunk size materially affect the ratio.

Reproduce with npm run bench. Set XMARKDOWN_CHUNK_SIZE to test a different delivery size. The benchmark source is run.js.