AT&T's Open-Source OTel-2.0 Beats Larger AI Models on Telecom Benchmarks

A coalition led by AT&T and the GSMA has released OTel-2.0-LLM-31B-IT, an open-source large language model post-trained for telecommunications workloads. Built on Google's Gemma 4 31B-IT and shipped under an Apache 2.0 license, it is the first release in the OTel 2.0 family and currently sits at the top of the Open Telco AI leaderboard. The weights have already crossed 6.8 million downloads on Hugging Face, a signal that operators and vendors are actively pulling domain-specific checkpoints into their stacks.

Why bother with a telecom-only LLM

Frontier models are strong at generic reasoning but weak on the specialized vocabulary that runs mobile networks. They still stumble on complex network topologies and vendor-specific telemetry unique to the sector. Telco-tuned models anchor the AI in the actual realities of network operations, letting the system interpret nuanced technical logs, diagnose performance bottlenecks, and reason about industry protocols with the precision real-time systems demand.

The leaderboard evidence backs the pitch. The top three performers on the Open Telco AI benchmarks are all domain-adapted models rather than general-purpose ones, with the new OTel 2.0 model on top. A 31B specialist is outrunning much larger generalists on standards interpretation and network Q&A.

The data pipeline behind the weights

OTel 2.0 starts from a curated corpus of telecom standards, then multiplies it through synthetic data generation. The model was post-trained on a 400 billion token telecom dataset filtered from more than 1 trillion processed tokens, incorporating a 15-billion-token base provided by the GSMA and compiled from seven major standards development organisations (SDOs): 3GPP, ETSI, GSMA, CAMARA, ITU, O-RAN, and TM Forum. Those 15B raw tokens were expanded past a trillion using Red Hat's open-source Synthetic Data Generation Hub, from which roughly 440B were sampled for post-training.