Telecom is one of the few domains where general-purpose AI models still fumble the basics: dense jargon, arcane protocols, and standards documents that read like legal code. Open Telco, or OTel, aims to fix that with an open resource built specifically for telecom tasks.
OTel bundles derived datasets for retrieval, reranking, instruction tuning, and safety/abstention, alongside 30 full-parameter post-trained baseline models spanning embedding, reranking, and language-model families. Post-training numbers: embedding retrieval hits 93.5% NDCG@10, reranking reaches 0.952 MRR@10, and language-model correctness reaches 88.2%. As of May 3, 2026, the released models had been downloaded more than 16 million times and picked up over 157 pieces of media coverage worldwide. The project builds on earlier open telecom datasets and benchmarks rather than starting from scratch.
Telecom operators have long been stuck choosing between generic LLMs that mangle domain terminology and expensive proprietary systems built in-house. A documented, reproducible baseline with held-out evaluation partitions gives smaller carriers and vendors a real starting point instead of reinventing retrieval and reranking pipelines from zero. That matters more as telecom networks lean on AI for tasks like fault diagnosis and network configuration, where wrong answers carry real operational cost.
Sixteen million downloads and 157 media hits are the kind of numbers that show up in a pitch deck as much as a paper. They measure interest, not whether OTel's models actually hold up inside a live network.