AT&T’s OTel 2.0 is now live: the largest and best performing open-source model built for telecoms

Now live and top of the leaderboard on Open-telco.ai, the new model is a post-trained version of Gemma 4 31B-IT built using 400 billion telecom-specific tokens selected from more than 1 trillion processed tokens. GSMA’s Louis Powell explores why this represents an important step towards delivering telco-grade AI for the telecommunications industry.

As AI becomes increasingly embedded in telecoms operations many challenges still remain. Namely, frontier models lack the specialised knowledge needed to understand telecom networks, standards and services. To help bridge this gap, AT&T has released OTel 2.0, the best and largest post-trained open model built specifically for the telecoms industry.

Why the industry needs its own models

General-purpose AI models have come a long way, but they weren’t built with telecoms in mind. Ask one to interpret an industry standard or troubleshoot a live network issue, and the cracks start to show, not because the models are not capable, but because the data they learned from barely touches this domain.

That gap shows up in the results. The top three performers on the Open Telco AI benchmarks are all domain-adapted models, not general-purpose ones, with the new OTel 2.0 model top of the Open Telco AI leaderboard. This clearly illustrates that domain adapted models are highly accurate and can be significantly smaller in size.

Just as healthcare, financial services and manufacturing are developing domain-specific AI approaches, the telecoms industry needs models trained on its unique standards, protocols and operating environments. This is not just about accuracy it is about enterprise requirements; reducing costs and maintaining control by deploying models across clouds and on-premise as needed. Accuracy and requirements together are vital for operators deploying telco-specific use cases, like network troubleshooting, product development, network configuration and more.

Building a corpus that didn’t exist

Since there’s no telecoms equivalent to Wikipedia, the industry has had to build training data from scratch.

The initial data set comprised approximately 15 billion tokens, which the GSMA provided, by collecting and processing documents from seven standards development organisations – 3GPP, ETSI, GSMA, CAMARA, ITU, O-RAN and TM Forum – and converting dense technical specifications into material suitable for model training. Alongside this, GSMA and Pleias released the Telco Corpus, the largest pre-training dataset built for the telecoms industry to date, spanning around 10 billion tokens.

That dataset was then combined with additional data from AT&T with collaboration from Red Hat, Dell, Microsoft Azure and AMD, to build the 400 billion token training-set. The result is the first release in the OTel 2.0 family.

What comes next

OTel 2.0 is a significant step, but it is one step on a much longer journey. The industry still needs models in a range of sizes, since not all telecoms use case calls for a large language model – efficiency matters as much as accuracy.

The OTel models are open for a reason: they are meant to be downloaded and fine-tuned further with even more data, and GSMA is looking forward to seeing how operators, vendors and researchers adapt them for their own use cases. The long-term ambition is to create an open ecosystem of telecom models, datasets and tooling that helps the industry build telco-grade AI.

If you are using OTel models and getting interesting results, please get in touch: GSMA would love to showcase your work with its members via channels including the MWC series of events.