OTel says it's been downloaded over 16 million times. Nobody involved has put their name on it.
An arXiv paper posted October 7, 2026 describes Open Telco (OTel), a package of datasets, benchmarks, and 30 pre-trained AI models aimed at telecom-specific tasks: retrieval, reranking, instruction following, and knowing when to refuse an answer. The models are post-trained versions of 10 embedding models, 3 rerankers, and 17 language models, tuned on telecom data and scored against held-out test sets. The paper reports strong results: 93.1% NDCG@10 for retrieval, 0.947 MRR@10 for reranking, and 87.8% correctness for the language models. It also claims the released models had passed 16 million downloads and 157 media mentions as of May 3, 2026, five months before the paper itself appeared.
That gap matters because the abstract we have lists no authors, no institution, and no sponsoring company. Telecom-specific AI is a real niche, and carriers do need models that parse network configs and spec documents without inventing details, so strong benchmark numbers are worth attention. But download counts and press-mention tallies are the kind of stats companies cite in funding decks, and there is no independent source here to confirm them.
Open-source AI projects usually earn trust through reproducibility and a visible team behind the numbers, and a toolkit downloaded millions of times that will not say whose lab built it is asking for credit it has not earned yet.