Published
Updated
A 100,000-line needle takes 18.0 minutes with Jev and 24.7 without
This topic was created 31 days ago, and the information it contains may have evolved or changed since then.
El Asombroso Nosequien posted elalber2000/jev-rlm, a fork of a minimal recursive-language-model loop. On one 100,000-line file both runs returned 2867825. results.json records 1,081 seconds with 82 Jev calls and 1,480 seconds with none. The post's dollar figures are an estimate. The file has no charge field.
El Asombroso Nosequien posted on September 23 a small recursive loop that looks for one sentence in a long file, with Jev on and with Jev off. The repo is elalber2000/jev-rlm, a fork of alexzhang13/rlm-minimal. The license file is MIT. The post says the Jev arm finished in about 27% less time, 24.7 minutes down to 18.0, and at about 43% lower estimated cost, $0.155 down to $0.089, off-peak.
main.py builds the file itself. Seed 42. The call site asks for 100,000 lines, not the function’s default of one million. A random seven-digit number is written into one line between 40% and 60% of the way through, and the query is “I’m looking for a magic number. What is it?” The root model and the recursive model are both deepseek/deepseek-v4-flash-0731, through OpenRouter. When the flag is on, the Jev model string is typesafe/jev-1.13. The loop stops after 10 iterations. The copy of use_jev on the branch is False.
data/results.json has both runs. With Jev: 10 root calls, 1 recursive call, 82 Jev calls, 70,055 characters into the root model, 8,188,252 characters into Jev, 1,081.05 seconds. Without Jev: 7 root calls, 78 recursive calls, no Jev calls, 153,118 characters into the root model, 3,975,146 characters into the recursive model, 1,480.50 seconds. Both results are the sentence “The magic number is 2867825.” Those clocks are the post’s 18.0 and 24.7 minutes. The difference is 399.45 seconds, 27.0% of the longer run.
The JSON has no dollar field. The $0.155 and $0.089 lines are the post’s estimate, and the post says off-peak. The file stores characters, not tokens, and this desk did not price them. The README tells you to read data/reslts.json. The file on the branch, and the path main.py writes, is data/results.json.
Both arms found the same number on the same seed. The Jev arm made 82 Jev calls and one recursive call. The other arm made 78 recursive calls. The results list stops there. We did not run the script.
This site's reading
Editorial notes evaluating claims against primary sources, contextualizing findings alongside related implementations, and defining technical terms.
Verify
Primary post is @sancheznoseke, display name El Asombroso Nosequien, September 23, 2026, 08:40 UTC. It says a needle in 100K lines took about 27% less time, 24.7 minutes to 18.0, and about 43% less estimated cost, $0.155 to $0.089, off-peak. Repo elalber2000/jev-rlm, MIT, forked from alexzhang13/rlm-minimal.
main.py builds 100,000 lines, seed 42, root and recursive model deepseek/deepseek-v4-flash-0731, jev_model typesafe/jev-1.13 when the flag is on, provider OpenRouter, max_iterations 10. The checked-in flag is False.
results.json, use_jev true: 10 root calls, 1 recursive call, 82 Jev calls, 8,188,252 Jev characters, 1,081.05 seconds, result 2867825. use_jev false: 7 root calls, 78 recursive calls, 0 Jev calls, 3,975,146 recursive characters, 1,480.50 seconds, same result. No charge field.
README names data/reslts.json. The file on the branch is data/results.json. We did not run main.py.
Compare
AstroHan's context filter is a different saving: 25 of 30 tasks against 22 of 30, and a token count in the hundreds of millions. This file is one needle, two runs, same seed, same answer. dejevu's flights comparison reprints a published Jev time and does not rerun Jev.
Here both arms are in the same JSON file.
Terms
- typesafe/jev-1.13
- The OpenRouter model string main.py passes when use_jev is true. The script does not print a served version such as jev-1.13.0. The checked-in default is use_jev = False. results.json contains one true run and one false run.
- 1,081 seconds
- The Jev arm's elapsed time in results.json, 1,081.05 seconds, which is the post's 18.0 minutes. The other arm is 1,480.50 seconds, the post's 24.7 minutes. The file records characters and call counts. It does not record the post's $0.155 and $0.089.