Linda-Pro 1.3.2 and Linda-Pro Lite: two local detectors of AI-generated text in English, Polish and Russian

Technical report on Linda-Pro 1.3.2, a local, privacy-preserving detector of AI-generated text in English, Polish and Russian that runs on the user's own computer without cloud services. The package, installed by desktop application version 2.0.3.7, now contains two tiers. Linda-Pro is the unchanged 1.3.1 ensemble (DeBERTa-v3-large and mDeBERTa-v3-base voters plus stylometry); its published measurements stand. Linda-Pro Lite is new: one distilled multilingual-e5-base student (12 layers, about 278 MB, fast fp16 storage in an fp32 ONNX graph) plus stylometry, trained to reproduce the ensemble's scores and built to run on an office computer processor without a graphics card. A 3,800-word text is checked in 1.2 s on four threads of a modern processor, against 226 s for the full tier on an office Ryzen 5 PRO 2400GE. On held-out sets (not used for training) Lite flags 68.2% of HumanizerBench texts (Linda-Pro: 81.0%), 74.1% of an October 2026 humanizer cycle (84.0%), 81.4% of texts by three unseen generators (95.5%), 96.3% of plain AI essays (99.8%), 51.6% after a stealth humanizer (88.0%), 75.4% of Polish AI texts (79.4%) and 74.6% of Russian AI texts (80.3%); false positives on human texts stay at 0.0-2.2%. Lite is weaker on deliberately disguised, short or older small-model texts; for important decisions use the full tier. Caveats and limitations (small Polish and Russian test sets, plus-minus 5 points; monthly changing humanizers; not an independent audit, not peer reviewed; reports of bugs found and fixed in packaging) are stated in full. Version 1.3.1 report: https://doi.org/10.5281/zenodo.23080472 . Version 1.0 report: https://doi.org/10.5281/zenodo.23072494 . Application and evaluation script: https://github.com/Lendarixon/Linda . Model weights: https://huggingface.co/Lindarixon/Linda-Pro and https://huggingface.co/Lindarixon/Linda-Pro-Lite . Self-published, not peer reviewed.