// nuwright
● active · built; not released

quench

Opens multi-gigabyte JSONL files on a Mac to read, filter and profile them.

job
open large datasets
domain
developer tools · training data
runs
local
class
tool
scale
tested on a 40 GB file of 24.9 million lines
since
2026

One call

  1. 01indexLine breaks found and indexed as the file streams in
  2. 02windowFirst rows on screen while the index grows
  3. 03schema5,000 lines sampled to infer the columns
  4. 04drawOnly the rows on screen are parsed
  5. 05profileFilters and a token histogram stream in the background

In · out

inputs
  • JSONL file
  • compressed .jsonl.gz file
  • a local tokenizer file
outputs
  • a table that scrolls millions of rows
  • conversation detail pane
  • token histogram
  • filtered rows

What is measured

Against: Its own earlier version (58 s to open 40 GB) and the disk's raw read speed (6.13 s).

Opens a 40 GB file in 6 s, down from 58 s and at the disk's own read speed; first rows on screen in under 3 msmeasured
Reopens the same file in under a secondmeasured
Scrolls 20 million rows at 60 frames a secondmeasured
Filters about 475,000 rows a second and searches about 1.1 million; first results in about 35 msmeasured
Under 500 MB of memory for a 40 GB filemeasured

Stack · links

RustGPUIrayonmemchrsimd-jsontokenizers
last updated 2026-09-23 · from source, not marketingclaims carry their own status