Telecom Operators Build Network Operations Around Open Models

Telecom operators are combining open and proprietary AI models for network operations, customer service and local-language applications, giving them more control over customization, data protection and deployment.

Open-source models and software are important to the AI strategies of 89% of surveyed telecom respondents. The approach is emerging as a hybrid model in which operators combine external systems with proprietary data, internal network expertise and governance.

NVIDIA announced the 30-billion-parameter Nemotron 3 Large Telco Model for tasks including network configuration and customer-incident triage. AdaptKey fine-tuned the model with open telecom datasets, while NVIDIA published a fine-tuning recipe using its NeMo libraries so operators can adapt models to their networks, customers and procedures.

SoftBank Corp. uses its in-house Sarashina model alongside external open models, including NVIDIA Nemotron, to develop its Large Telecom Model. The company built a synthetic-data pipeline for the model on March 17, 2026, using NVIDIA NeMo Safe Synthesizer and differential privacy to protect confidential network information.

AT&T post-trained OTel 2.0 with more than 400 billion tokens on AMD GPUs and processed more than 1 trillion tokens for training through Microsoft Foundry. Its model strategy combines proprietary and open-source technologies for production workloads that require scale, reliability and governance.

Sahabat-AI expanded on June 2, 2025, when Indosat Ooredoo Hutchison and GoTo introduced a 70-billion-parameter model with a multilingual chat service. The system supports Bahasa Indonesia, Sundanese, Javanese, Balinese and Batak, while its data is stored locally in Indonesia.

Open models give telecom operators access to development assets that can be adapted for specialized workloads. The cited projects focus on network automation, incident handling, data protection, local-language services and deployment on private infrastructure.