News
AT&T releases OTel 2.0, an open AI model built for telecom networks
AT&T’s OTel 2.0 brings a telecom-specific open model to the Open Telco AI leaderboard, trained on standards and industry data for network-focused work.
AT&T has released OTel 2.0, an open AI model designed specifically for telecommunications. The announcement matters because operators are asking general-purpose models to interpret standards, diagnose network problems and support configuration work that depends on specialist vocabulary and tightly controlled procedures. The GSMA says OTel 2.0 is now at the top of the Open Telco AI leaderboard, giving the industry a new reference point for domain-focused models.
OTel 2.0 is a post-trained version of Gemma 4 31B-IT. Its training set contains 400 billion telecom-specific tokens selected from more than one trillion processed tokens. That scale is important, but the more consequential detail is the source material: the model was built around the standards, protocols and operating language that shape real telecommunications work.
Why a specialist model is useful
Large general-purpose models can explain familiar concepts, write code and summarize documents, but they do not automatically understand the context of a mobile network. A request about a radio-access procedure, a standards clause or a live fault can require knowledge that is sparse in ordinary web data. A confident answer that misses one technical constraint is not useful to an operator, and can be actively risky if it is treated as an instruction.
According to the GSMA announcement, the three highest performers on the Open Telco AI benchmarks are domain-adapted models rather than general-purpose systems. The GSMA presents that result as evidence that a smaller, specialised model can be highly accurate for a defined industry while remaining easier to control than a much larger model trained for every subject.
That distinction also affects deployment. Telecom operators may need to run models across public clouds, private infrastructure or on-premises systems, depending on the sensitivity of the workload and the network environment. A focused model can reduce the amount of irrelevant knowledge an organisation must manage, while making it easier to evaluate performance against the tasks that actually matter: network troubleshooting, product development, configuration support and technical documentation.
Training data built for telecoms
The project began with a problem that is easy to overlook: telecommunications does not have one universally complete, plain-language knowledge base equivalent to Wikipedia. Technical knowledge is distributed across specifications, implementation guidance, industry documentation and operational material. The GSMA says it created an initial dataset of about 15 billion tokens from documents produced by seven standards organisations: 3GPP, ETSI, GSMA, CAMARA, ITU, O-RAN and TM Forum.
Those documents were processed and converted into material suitable for model training. The GSMA also worked with Pleias on the Telco Corpus, described as a pre-training dataset for the telecoms industry containing about 10 billion tokens. Additional data from AT&T was then combined with collaboration from Red Hat, Dell, Microsoft Azure and AMD to form the 400-billion-token training set behind OTel 2.0.
This provenance gives the model a clearer technical centre of gravity than a generic chatbot. It does not make every answer correct, and it does not remove the need for engineers to verify outputs against current standards, network policies and local change-control processes. It does, however, create a more relevant starting point for experiments in which telecom-specific terminology and relationships are central to the task.
Open by design, but not finished
OTel 2.0 is presented as the first release in a wider OTel 2.0 family. The GSMA says the models are open so operators, vendors and researchers can download them and fine-tune them with additional data. That approach could let organisations adapt a base model to a particular network domain, language, operational workflow or internal documentation set without starting from zero.
The project is also an acknowledgement that one model will not fit every telecom use case. A troubleshooting assistant, a standards-search tool and a configuration copilot may need different balances between accuracy, latency, memory use and explainability. The GSMA explicitly points to the need for models in a range of sizes, because efficiency can matter as much as raw capability when inference runs close to network operations.
For mobile users, the release is unlikely to appear as a new phone feature overnight. Its nearer-term impact is behind the scenes, where specialised models could help operators improve service assurance, automate repetitive engineering work or build more capable support tools. Any customer-facing deployment would still depend on testing, access controls, privacy safeguards and human approval for actions that can affect connectivity.
OTel 2.0 therefore looks less like a finished consumer product than an important infrastructure experiment. Its value will be measured by how reliably the open model performs on real telecom tasks, how transparently those results can be checked and whether the surrounding ecosystem produces smaller, safer and more efficient versions. By releasing both a model and a data-building direction, AT&T and the GSMA are making a concrete case for AI that understands the networks it is meant to help operate.