Lindarixon/Linda-Pro-Lite

Linda Loupe: a small, fast AI-text detector (English, Polish, Russian)

Windows application 2.0.6.2 — 17 October 2026

Download the release · Release notes · Try the online demo

Clear warnings before analysis, safer document handling and model-folder moves, responsive Loupe cancellation and redacted diagnostic metadata. Prefer at least 300 words of continuous prose; lists, forms and slide text may be unsuitable. Model package 1.3.4, tokenizers and thresholds are unchanged; short AI inserts can still be missed and results do not prove authorship.

The hosted demo sends text to Modal servers: 12 successful checks with a shared-network daily limit, up to 6,000 characters, TXT/MD/DOCX. Do not submit confidential text. The desktop app processes text locally and has no demo quota.

Formerly named Linda-Pro Lite.

Linda Loupe is the compact version of Linda Assay. It is one distilled model (multilingual-e5-base, 12 layers, ONNX) plus stylometry that together reproduce the large ensemble's verdicts closely enough for everyday use, and it runs on the CPU of any ordinary computer in seconds. No graphics card is needed. The same Windows app (github.com/Lendarixon/Linda) can run with only Lite, only Linda Assay, or both.

Two versions: Linda Assay and Linda Loupe

The same app, two sets of models. Install either one or both; the installer asks, and Settings let you change it later (the app also works with only one of them).

• Linda Loupe: a small, fast model for any computer, including old laptops without a graphics card. Slightly less accurate.
• Linda Assay: the full ensemble, the most accurate; best with a modern graphics card. Not sure? Take Lite: it runs everywhere, and Linda Assay can be added later.

	Linda Assay	Linda Loupe
Texts rewritten by 16 "humanizers" (HumanizerBench), found	81.0%	69.4%
Newer humanizer texts (Oct 2026), found	84.0%	75.8%
Texts of three unseen AI generators, found	95.5%	83.0%
Polish AI texts found (false alarms)	79.4% (1.4%)	72.2% (1.5%)
Russian AI texts found (false alarms)	80.3% (0.6%)	77.3% (1.2%)
False alarms, non-native exam essays (TOEFL)	5.5%	2.2%
False alarms, human texts of a mixed set (MAGE)	2.0%	0.4%
Check of a 3,800-word text	1-3 s with a modern graphics card; minutes on a processor alone	about 1-5 s on one ordinary processor (measured 1.2 s on a Ryzen 7, 4 threads)
Download	about 2.9 GB	about 0.3 GB
Graphics card	recommended	not needed

Recommended computer:

	Linda Loupe	Linda Assay
System	Windows 10 / 11, 64-bit	Windows 10 / 11, 64-bit
Processor	any 4-core processor from about 2017	4 cores or more
Memory (RAM)	4 GB (8 GB is comfortable)	8 GB (16 GB recommended)
Graphics card	not needed	DirectX 12 card with 4 GB of video memory or more; works without it, but slowly
Free disk space	2 GB	6 GB (an SSD is better)

Our own measurements on public test sets (English, Polish, Russian), not an independent audit; one calibration rule for both versions. The full Linda Assay needed 226 s for a 3,800-word text on an office Ryzen 5 PRO 2400GE; Lite is expected to take a few seconds there (not measured on that processor yet). Lite weights have their own repository: Lindarixon/Linda-Pro-Lite. From Python: python download_models.py --set lite, then python -m linda_pro file.txt --lite.

Use it

pip install -r requirements.txt
python -m linda_pro my_text.txt --windows        # verdict + per-window scores; the language is detected automatically

from linda_pro import load_detector
print(load_detector("sensitive").detect([text])[0]["verdict"])   # "ai" | "uncertain" | "human"

Modes: sensitive (default) and precise (fewer false accusations, for schools and universities). A verdict is a screening aid, never the sole basis for a decision about a person.

How it was made

• Teacher: the Linda Assay 1.3.1 ensemble (Essay-M + Multi-M, stylometry).
• Student: multilingual-e5-base (MIT), 12 layers, vocabulary pruned to the tokens of English, Polish and Russian; trained to reproduce the teacher's window scores and its pooled representations on about 149,000 windows of public human texts and texts of open-weight models, plus Polish and Russian texts written by T-lite and Bielik and rewrites by Qwen3-8B in "humanizer" styles. Texts generated by Llama models are excluded (rows whose generator or source names a Llama or Nemotron model were removed, together with texts derived from them).
• Decision rule: student 0.8 + stylometry 0.2 (Polish and Russian: 0.8 / 0.1 stylometry / 0.1 per-language stylometry); thresholds are quantiles of development human texts and were not tuned on the check sets.
• Data sources and licences: see DATA_BOM.md and NOTICE.md (same lineage as Linda Assay 1.3.1 plus the generated texts above).

Limits

Lite finds fewer texts after "humanizer" rewriting than Linda Assay; on texts under the recommended 300 words results are less reliable; Polish and Russian test sets are small (350 + 350 texts; about +-5 points). Not an independent audit.

Licence

Free for personal non-commercial use (CC BY-NC-4.0); commercial use needs a paid licence (COMMERCIAL_LICENSE.md, lindapro.support@proton.me).

Citation

If you use Linda Assay or Linda Loupe, please cite the technical report for version 1.3.2 (Zenodo, DOI 10.5281/zenodo.23261670). Earlier reports: version 1.3.1, 10.5281/zenodo.23242014; version 1.1, 10.5281/zenodo.23080472; version 1.0, 10.5281/zenodo.23072494.

@techreport{manzyuk2026lindapro132,
  title     = {Linda Assay 1.3.2 and Linda Loupe: two local detectors of AI-generated text in English, Polish and Russian},
  author    = {Manzyuk, Vladyslav},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.23261670}
}

Documentation alignment — 17 October 2026

Current Windows app: 2.0.6.2; the combined update feed provides model package 1.3.4. This date describes the application/feed, not a new release of every individual model in this repository. See trust, benchmark brief and root LICENSING.md. Historical metrics retain their original versions and denominators.

Current deployment note: Windows Assay 2.0.6.2 uses FP32 stable DirectML (8 GB VRAM recommended; 4 GB unvalidated), Loupe uses CPU. Desktop history/update/licence behavior differs from the CLI and hosted demos. Historical speed and accuracy values must retain their original versions, modes and populations.

Free 14-day commercial evaluation

Free commercial evaluation lasts 14 calendar days from first launch, without a paid key, registration or automatic billing. This is a licence condition; the app does not enforce a timer. Continued commercial use then needs a paid licence. Personal non-commercial use stays free without a time limit. See COMMERCIAL_LICENSE.md for the full grant.

Stylizer

The online demo has a Stylize button (rewrites a whole text, AI-looking sentences or your selection; it does not hide authorship). Adapters: Lindarixon/Linda-Stylizer (CC BY-NC-4.0, same terms as the detectors; meaning is not verified automatically).